My ComfyUI workflows for using LTX 2.3
This workflows are used by me to create my art.
They are optimized created of my latest knowledge to enhance the outcome.
Versions & Information👇👇👇👇👇👇👇👇
👉 Please read below and the file descriptions "About this version" for more info's.
⚠️ Do not use the workflows with the "Nodes 2.0 beta" from ComfyUi or it will mess up things.
👇👇👇👇👇👇👇👇
What you get from the comfy workflows:
♨️ Easy controls
✅ As less as possible dependencies
🪧 Detailed documentation
⛓️ Highly automatic logic
✨ Optimized results
🔖 Bookmark-Shortcuts with number keys
Types of workflows
OmniForge C-LTX23
🎥 I2V, FLF2V, T2V, V2V + Audio
🤝 Video resolution matching - Fully automatic scaling
⚙️ Chunking (Seconds/Frame based)
🎞️ LTX Latent Upscaler (2x native latent)
🪫 Low VRAM optimizations
✨ Multiple Resolution Upscalers
🔗 VAE options
Full VAE or Tiled VAE
🔊 Multiple Audio options
🫥 Watermark option
📢 Soundmark option
🧮 Color transfer feature
🔖 Bookmark-Shortcuts - with number keys
🩻 Known issues and advice's
⚠️ Some workflows may set on webp av1 encoding (VHS node) - If your computer/setup missing drivers use any other like H265 or H264!
Install ffmpeg!
Update Comfyui and custom_nodes!
Update pytorch 2.9+cu128 or higher
Make sure to read where files/models should be placed inside the workflow
Check if the filepath for model/clip/vae match your system like Linux/Windows
Some Custom-Node-Packs need manual installation (e.g. RTX Node)
The plugin ComfyUI-DD-Translation can break node connection (avoid)
All older Versions are available inside my GitHub Repo.
Learned a lot from RuneXX LTX Workflows! They are awesome! Check them out! 🫶
Special thanks to @DustyDrab for helping with tests and bug-fixing!
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Description
Latest Changes / Fixes
TritonVAE (patcher) for massive speed-up
Added DMD Distilled lora suggestion
Added Torch enabled simple upscaler
Added Advanced Video Combine Node
Removed deprecated Frame Extraction
UX fixes
Scheduler optimizations
⚠️ Make sure to update to my latest "nightly" DaSiWa Node Pack
FAQ
Comments (533)
Showing latest 247 of 533.
For the V2V function are there other inputs or settings required other that video and audio upload? The workflow takes both inputs and runs without error, but does not produce an output video
For v2v minimum is a video file, extra audio is optional
@darksidewalker Hmmm the video preview generates below the main settings but no actual output is created; since I am using a video file and extra audio file, I was hoping I was just missing something else - using your current Comfy Installer - I will keep playing with it to see if I can figure it out - thank you!
@darksidewalker It looks like the models\frame_interpolation\rife_v4.26.safetensors was an abnormally small size, so I replaced it with https://huggingface.co/Comfy-Org/frame_interpolation/blob/main/frame_interpolation/rife_v4.26.safetensors
After that everything is working - I am not sure if maybe I just got interrupted or something went bad when I installed or if the file from your custom Comfy installer got corrupted, but you might want to take a peek at it
@FirstPrinciples That's is the same link my installer should dl with 22.7 MB
You are right, there was a bug in the dl url, downloading the webpage instead of the actual file. Fixed.
anyone able to use this with the newly released GGUF of Eros?
https://huggingface.co/vantagewithai/LTX2.3-10Eros-GGUF/tree/main
I keep getting blurred output. i didn't change any setting, i am getting normal output with the base LTX 2.3 GGUF from
i tried the base sulphur GGUF too and i get the same blurry videos output
https://huggingface.co/vantagewithai/Sulphur-2-Base-GGUF/tree/main
The distilled sulphur versions can get good outputs but its kinda random, I don't know much about all the different settings that could be changed to solve that. However the sulphur dev/base and 10eros model requires the distilled loras and if you read the civitai page about it there is an issue with a bunch of them that cause the blurring/colour flickering, they've posted the loras that work best but you need to have them at different weights for each pass which you can't do on this workflow without changing things which again I don't know how to do lmao if you use the distilled loras at 0.5 or lower the problem is minimal, unless you aren't using the loras at all and just getting blurry outpots like that anyway lol
When using the i2v feature, the character’s face is transformed into a more Western appearance.
At first, I thought the issue might be caused by the model or LoRA. However, when I tested the same model and LoRA with the default LTX 2.3 workflow for i2v, the character’s face did not change.
Could you please tell me how to prevent the character in the image from being altered into a Western-looking face?
I cannot tell, since when I use it is completely preserves the style of the input image.
Im having the same problems. Im testing around but nothing has worked so far
Can't find in the Omniforge workflow
FrameInterpolationModelLoader
in subgraph 'Settings and Backend'
and
FrameInterpolate
in subgraph 'Settings and Backend'
But I don't know what pack these are installed with and neither does the manager. Can anyone help point me to where those are?
Update comfyui
@Darksidewalker Yeah that was the first thing I tried. I must have some conflict with other nodes or something. Thanks for the reply and all that you do with models and workflows.
@EricRollei21 It is just a native comfyui node from comfyui 0.20.+
I'd like to be able to preview video samples captured in the first stage of the workflow to avoid wasting time generating imperfect videos. Once the first stage is complete, I won't be able to see the video quality if I'm doing other things and not continuously monitoring it. I hope you can optimize this.
I have no clue what you are talking about. 🤔 There is a preview.
@darksidewalkerAfter multi-frame stage 1 is completed, the screen freezes completely and cannot be previewed. I'm not sure if this is a problem with my COMFYUI.
Thanks much for the workflow!
Consider adding an "audio_start" option like the one in the "Video Generation (LTX-2.3)" node please.
I'm attempting to create a music video with a song, and therefore I put the entire 3 minute song in for the audio file but need to do it in 5-10 second bursts.
So if I have an option to start audio at X time in the audio I uploaded, it'd be do-able.
The basic video_ltx2_3_ia2v workflow has such an option. I'd much appreciate it in this workflow! :)
I can look into 😺
@darksidewalker o7 <3
I think the LTX workflow really needs the "Prompt Relay Encode (Timeline)" prompt node, which allows for precise segmentation control and I think it's very helpful for video generation.
Everything works pretty well, but how do I batch a set of images without going 1 by 1? Or set up prompt per set? I did try to load the multi image loader from LTX Director, but it didn't work that way. Can anyone help?
support orignal fp16 version?
Thank you for this amazing workflow :)
V2V function not working:
First of all huge thank you for your efforts to provide quality models and workflows, yours are the top notch..
I like this workflow very much, the "image to video" function works great, however i don't seem to have the same success with "video to video"
I always get "Still" videos with almost no motion, as i understand, this function should add sound to the uploaded video and extends it, if i tried to upload a 5 seconds video from "wan 2.2" without sound, the V2V function adds nice sound successfully, but the video is a 5 seconds still image! if i chose the 10 seconds or even 20 seconds options, it still a video with so little motion with no relation to the original uploaded video!
I tried different video output options but no luck, strangely enough, the v2v function works for me with a simpler workflow here:
https://civitai.red/models/2533889/ltxadd-v2v-add-audio-and-extend-any-video-ltx23
but i never managed to get it to work in your workflow and i really wish i realize the issue at my end
I did v2v just 2 days ago, so I know it is working. but I cannot guess what is wrong. May you post your settings or maybe screenshots on my discord?
is this compatible with the eros / sulphur model?
Sure
Hi, I have TypeError: SwitchNode.execute() missing 1 required positional argument: 'on_true' when color match is enabled
Apart from that problem, everything is working fine :)
Why does my settings and backend node look different from the preview? It just shows lots of 'DaSiWa Node Status Switch' toggles with true/false options.
may i ask what Chunking is? aside from making it faster, does it degrade video quality?
Chunking is described in the workflow nodes, it is not making things faster, even slower. Much slower, but saves VRAM.
It can actually help with retaining details and doing lipsync over longer videos; instead of processing the entire clip which can lead to temporal wandering (the character's lips just stop moving like 6 seconds in for example) chunking breaks it up into smaller segments then blends/overlaps them for a final product. I have found to be useful if making anything > 8 seconds
@FirstPrinciples nice hint
I've had great success using your workflows, especially the FMLF. It would be awesome if you could include the new Prompt Relay nodes in a future OmniForge workflow iteration. Thanks again!!
Hey dude, I saw your MiniMeme feature - Create small gifs in workflow, but I don't understand how to do it?
activate it, and have a small gif vid to your vid
I've been getting this persistent error ComfyUI-LTXVideo (1) using your setup tool but haven't been able to find any working solution to it, is there something I'm missing?
The LTX custom nodes installs a wrong kornia version, they have to fix it or you have to downgrade kornia to like 0.8.2 or earlier
Or run the setup tool again, pinned a working version
I solved this by downloading from Github an older version of the ComfyUI-LTXVideo costum nodes. The one of February 2026 if I remember correctly
Anyone else getting a watermark at the last frame even when watermark is turned off? It's illegible purple text with a yellow border.
Never got this can you post an example on discord? Preferably SFW
Am I the only one having lip-sync issues running FMLF on v3.6? I ran and continue to run v2.5 FMLF with 20s renders on 4090 with no issues. When I try to run the 20s FMLF using the v3.6, the quality is awesome but the lip-sync is way off and not sure why. Why does v2.5 run great but v3.6 does not, any explanation would be greatly appreciated. I prefer v3.6 quality but lip-sync isn't working for me.
It's the same for me
Oh, no, I figured it out. The "First-Frame Anti-Burn" function is to blame, it creates a small shift, you just need to disable it.
It does?
@Darksidewalker Yes, I don't know about other people, but that's the way it is with me.
@aurau673 possible, I may look into this, since it only cuts frames not audio
I also confirm that if you disable the first frame anti-burn option, the render will lip-sync as normal with no issues.
@Darksidewalker I also noticed that if you use the First-Frame Anti-Burn plus the Interpolation function the audio will be cutted at the exact half of the video.
@aurau673 omg thank you. i was losing my mind trying to figure this out.
I removed the interpolation for now, since it is not working with audio as intended.
The new V2V is AMAZING
original Wan generation and LTX V2V audio for comparison: https://civitai.red/posts/28746519
No issues here! Tested out multiple iterations of I2V and T2V, works great so far! Thanks for the workflow - seems to work a lot better than other LTX2.3 WFs I've tried
ComfySwitchNode has issue and workflow not working properly, I disabled all of them and now its working, you need to make them seperate logic, and all things in the same workflow is looking messy.
What?
Do you know why when I use FLF2V it cuts off about 0.5 seconds at the end of every video I generate? I'd like to create a looping video with the same two images, but I can't get it to loop perfectly because, for some reason, the last 0.5 seconds are cut off. I tested in the RuneXX FLF2V workflow and it works well so I don't know why it's not working in yours.
First time I hear this, don't really know, I don't do much flf2v, maybe it's a bug or setting
Great workflows as always; apologies if I'm missing something, but is the V2V option supposed to be affected by the video resize options (precision, resolution, etc...) that are selected? For some reason, any v2v I generate ends up coming out at a much higher resolution than the original video and it doesn't seem to change no matter what size option I pick.
Much appreciated.
Yes it is effected by the resolution nodes, if you do not want a change, you may select to not scale.
Getting "AttributeError: module 'sageattention._qattn_sm89' has no attribute 'qk_int8_sv_f8_accum_f32_fuse_v_scale_attn'" on kjnodes/ltxv_nodes for some reason with sage attention (2.2.0) enabled on RTX 5090. Any known fix?
I have a RTX5090 too and no error, maybe you selected a wrong sage option?
Patch Sage Attention KJ节点后面的LTX2 Mem Eff Sage Attention Patch节点忽略掉，这节点导致报错，因为这个LTX2 Mem Eff Sage Attention Patch节点需要最新的sageattention。
The LTX2 Mem Eff Sage Attention Patch node following the Patch Sage Attention KJ node is ignored, causing an error because this LTX2 Mem Eff Sage Attention Patch node requires the latest sage attention.
@Darksidewalker The issue might be with the LTX2 Mem Eff Sage Attention Patch node.
@ollin so sage2 not properly installed than?
@Darksidewalker I'm using latest 2nd generation SageAttention. I also made a barebones workflow with SageAttention to test it in isolation, and SageAttention worked there without any issues, so faulty installation of SageAttention is likely not the case.
@ollin Can confirm, removing the LTX2 mem eff sage attention node makes it work
@_Y_ The nodes work on my side, so something must going on with dependencies/incompatible settings, imho.
LTX2 mem eff sage attention needs latest sage 2 and triton kernels.
I'm on latest comfy, drivers and packages and all is working.
I tried on Win11 and Linux with my installer, so I assume something is missing on your end.
@Darksidewalker If may I ask, how big are your VRAM savings/speedup with SageAttention compared to with it off? I'm just not sure if it's worth to try and fix it.
Also, is the VRAM usage increasing by additional ~2GB during the second pass (where it does 3 steps) an expected thing? Just wondering at this point if there's also other things that may be broken in my case
@_Y_ sage2 brings like 5-10% more speedup, no VRAM savings.
"Hi! I'm getting a Windows fatal exception: access violation crash when trying to run the WAN 2.2 workflow. It crashes at load_diffusion_model in storage.py when loading the model file. This started happening randomly and now happens consistently even on a fresh restart of ComfyUI and Shadow PC. My setup: RTX A4500 20GB VRAM, 28GB RAM, PyTorch 2.9.1 cu130, ComfyUI 0.22.3. I also have LTX 2.3 workflow installed alongside WAN 2.2. Has anyone else experienced this and is there a fix? Thank you!"
access violations are OS side errors, you are rather has corrupted files, faulty drives or insuffizient permissions
@Darksidewalker I don't know what's wrong. I'm not very tech-savvy, but mine keeps crashing every time I run it. I don't know if it needs more VRAM or if I'm doing something wrong, but it's crashing every time now.
@jinmoriad most likely your OS is damaged or your drive is faulty
@Darksidewalker Love your work as always 🖤
Wanted to ask — how much paging file and free disk space do people generally need for LTX 2.3 to run smoothly?
And with V3 I'm getting some face blurring/ghosting — any tips to fix that?
Thanks for everything! 😊
RuntimeError: split_with_sizes expects split_sizes to sum exactly to 7680 (input tensor's size at dimension 2), but got split_sizes=[4096, 2048]
No clue why this happens to you.
Whatever you changed from 3.6 to 3.7 is the problem, because 3.6 works for me (exept first frames anti burn audio issue) but 3.7 not at all. I updated ComfyUI and all the nodes - same error. The error message comes after allow_fp16_accumulation = True and after the model initializes, when the first computing step out of 12 starts.
Ki says this:
The error you are seeing happens because you are trying to run the newer LTX-Video 2.3 model inside a legacy LTX-Video 2.0 or traditional video pipeline workflow.
Unlike older versions, LTX 2.3 introduces an AVTransformer3DModel architecture. This native architecture handles Audio + Video simultaneously inside an integrated Audio-Video (AV) latent space.
When ComfyUI runs the sampler, the underlying PyTorch code tries to split your incoming text conditioning context tensor into separate video and audio conditioning chunks (v_context_dim=4096 and a_context_dim=2048). It expects the incoming tensor size to match those combined dimensions. Because a legacy text encoder setup was used, it passed a tensor size of 7680 instead, causing PyTorch's split_with_sizes function to throw a mismatch exception and crash.
How to Fix It
To clear this error, you need to reconstruct your workflow to accommodate the specialized Audio-Video pipeline requirements of LTX 2.3.
1. Update Your Custom Nodes
Your traceback points directly to comfyui-kjnodes (ltxv_nodes.py). Open your ComfyUI Manager and make sure ComfyUI-KJNodes and ComfyUI-LTXVideo are updated to their latest versions to ensure all the updated node configurations are present.
2. Use the Dedicated Text Encoder Loader
Standard CLIP loaders or older text encoders will pass wrong tensor shapes. LTX 2.3 expects the Gemma 3 12B IT text encoder. You must load it using the dedicated LTXAVTextEncoderLoader node, which correctly formats the multimodal embedding sizes.
3. Handle Audio-Video Latent Spaces
Instead of passing a standard latent image straight into the sampler, you must multiplex the video and audio latents together before sampling:
LTXVEmptyLatentAudio: Add this node to generate an empty audio latent space.
LTXVConcatAVLatent: Take your empty video latent and your empty audio latent, plug them both into this node, and feed this combined output into your sampler.
4. Separate the Latents Before Decoding
Once the sampler finishes its job, the output is still an AV-combined tensor. You cannot pipe it straight into a standard VAE Decode node:
LTXVSeparateAVLatent: Route the output of your sampler into this node first. It will split the combined tensor back into clean, individual video and audio streams.
You can then safely connect the separated video latent out to your VAEDecode node.
@vicaut Well tour AI is wrong, except you changed my nodes with incompatible custom nodes
Make a new folder, use my installer and test this. Tell me if it still not work.
Hi,
First, thank you for providing the DaSiWa LTX23 workflow. I have been testing it and encountered an issue I would like to ask about.
When the workflow is set to T2V mode, the node "Load Video FFmpeg (Upload) 🎥🅥🅗🅢" still attempts to access or require a video source. Because T2V should not need any video input, this behavior seems unexpected.
In my case, the workflow tries to read a dummy.mp4 file, which leads to an FFmpeg error because no actual video input should be required in T2V mode.
My question is:
Is this behavior intentional due to the unified workflow design (I2V / FLF2V / T2V / V2V / Audio combined), or is it a bug?
If it is intentional, is there a recommended way to disable or bypass the video input nodes when using T2V mode only?
Any guidance would be appreciated. Thank you for your time and for your work on these workflows.
Best regards
Unfortunately this is a limitation due to the comfyui design. Thats because I also placed dummy files with my installer.
I did not found another way to provide multi-functionality without that design.
@Darksidewalker Thanks for the quick reply. Now I understand the limitation — that was very helpful. I appreciate your work and support.
You can also bypass the section, if not needed.
thanks for your work, but I got errors, shown as following. I just tested the text to video on default changed resolution, and bypass the video input (since the `placeholder.mp4` issue). I disabled the lightspeed and switch on text to veido. All models are downloaded from the links in the workflow, the comfy is latest (I run comfy update before importing the workflow)
[ERROR] !!! Exception during processing !!! The size of tensor a (2299) must match the size of tensor b (313728) at non-singleton dimension 2 [ERROR] Traceback (most recent call last): File "/home/user/comfy/ComfyUI/execution.py", line 536, in execute output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/execution.py", line 336, in get_output_data return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/custom_nodes/comfyui-lora-manager/py/metadata_collector/metadata_hook.py", line 171, in async_map_node_over_list_with_metadata results = await original_map_node_over_list( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/execution.py", line 310, in _async_map_node_over_list await process_inputs(input_dict, i) File "/home/user/comfy/ComfyUI/execution.py", line 298, in process_inputs result = f(**inputs) ^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy_api/internal/__init__.py", line 149, in wrapped_func return method(locked_class, **inputs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy_api/latest/_io.py", line 1852, in EXECUTE_NORMALIZED to_return = cls.execute(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy_extras/nodes_custom_sampler.py", line 967, in execute samples = guider.sample(noise.generate_noise(latent), latent_image, sampler, sigmas, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise.seed) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/samplers.py", line 1319, in sample output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/patcher_extension.py", line 114, in execute return self.wrappers[self.idx](self, *args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/custom_nodes/comfyui-kjnodes/nodes/ltxv_nodes.py", line 916, in __call__ out = executor(noise, latent_image, sampler, sigmas, denoise_mask, combined_callback, disable_pbar, seed, latent_shapes=latent_shapes) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/patcher_extension.py", line 106, in __call__ return new_executor.execute(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/patcher_extension.py", line 113, in execute return self.original(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/samplers.py", line 1257, in outer_sample output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/samplers.py", line 1232, in inner_sample samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/patcher_extension.py", line 113, in execute return self.original(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/custom_nodes/comfyui_smznodes/smZNodes.py", line 114, in KSAMPLER_sample return orig_fn(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/samplers.py", line 995, in sample noise = model_wrap.inner_model.model_sampling.noise_scaling(sigmas[0], noise, latent_image, self.max_denoise(model_wrap, sigmas)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/user/comfy/ComfyUI/comfy/model_sampling.py", line 97, in noise_scaling return sigma * (s * noise) + (1.0 - sigma) * latent_image ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ RuntimeError: The size of tensor a (2299) must match the size of tensor b (313728) at non-singleton dimension 2There is no way I can know what that come from, sorry
Boss, I don't understand why the input image size is 1056×960, with settings precision_presets:1.2mp-hd++, resolution_presets:1080p, scale_from_image:yes, using the goldenlaceV3-nvf4 format model, but the output video is very blurry. I'm a beginner in video generation.
I probably know the reason now: ltx-2.3-22b-distilled-lora is not turned on.
golden lace v3 has no distillation baked in, right
I'm wondering if anyone else is having audio issues...I'm getting nothing but a "hot mic" sound in the Model Audio mode, and when I use any kind of audio input it just throws that in as a soundtrack, but doesn't use it to guide the resultant output in any way. I have another LTX workflow that functions perfectly, so I took all of the models from this workflow and popped them into that workflow and the other workflow still works perfectly, so the issue isn't model related. I've turned off everything to try and just do a basic I2V test with model audio (no upscaling, no interpolation) with the same results. I threw in an audio decode and audio preview node after the first pass and the "hot mic" audio exists there, so the issue is upstream of the first pass sampler. Any ideas?
What is the other workflow - i'm interested :D
I did like 20 vids today, but did not notice that issues you describe, any references to that so I can have a look?
@Darksidewalker Yeah, I'm just getting started with LTX so I'm assuming it's something specific to my setup, but I've tried as many different tweaks as I can think of and I'm getting the same result every single time. ChatGPT wants me to gut your workflow to make it work like the other workflow that I do have working but your workflow is so well designed and outputs such spectacular video that I don't want to mess it up. What specifically do you need and what is the best way for me to get it to you?
@Darksidewalker Yeah, I'm just getting started with LTX so I'm assuming it's something specific to my setup, but I've tried as many different tweaks as I can think of and I'm getting the same result every single time. ChatGPT wants me to gut your workflow to make it work like the other workflow that I do have working but your workflow is so well designed and outputs such spectacular video that I don't want to mess it up. What specifically do you need and what is the best way for me to get it to you?
@vicaut The one I have working is the ltx23facefixWorkflow_v10 from another contributor on this site. That one outputs video with synchronized audio, but the initial video quality isn't great and the default settings seem to take the initial input image as more of a suggestion than anything. I'm sure with some tweaking I can get it to output something closer to what I'm looking for, but right now I'm just testing out LTX.
@Darksidewalker So I figured it out (well, my GPT figured it out anyway). I had a patch applied to fix another issue from an earlier graph that was padding my latent audio noise with zeros in order to force it into the correct shape...this messed up your sampler node. Not really sure why the other graph was working and yours wasn't, but I'm getting great results now. Thanks for your great work!
DaSiWa, everything you create is amazing, thanks for the hard work!
i have one question: is there a possibility to select multiple loras for this workflow like in your WAN 2.2 workflow?
sure just throw them into the lora loader
@Darksidewalker Your ComfyUI installer is also godlike, i always had problems with the recent updates, but with your easy to use installer everythin works like a charm :)
@HorstBorst29 Glad it was useful!
I have error Loop (2277,2205) - not submitting workflows
I suggest you to use LTX Director:
https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
You can use not just first, middle and last frame - but as many frames you like. You can even make frames before the first image.
# ComfyUI Error Report ## Error Details - Node ID: 2277:2723 - Node Type: LTXVConcatAVLatent - Exception Type: TypeError - Exception Message: TypeError: 'NoneType' object is not iterable ## Stack Trace ``` File "J:\ComfyUI_windows_portable\ComfyUI\execution.py", line 535, in execute output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
Boss, there is an error in the color matching function node, and there is a lack of reference input images.
TypeError: ColorMatchV2.execute() missing 1 required positional argument: 'image_ref'
# ComfyUI Error Report
## Error Details
- Node ID: 2277:3090
- Node Type: ColorMatchV2
- Exception Type: TypeError
- Exception Message: TypeError: ColorMatchV2.execute() missing 1 required positional argument: 'image_ref'
## Stack Trace
```
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 535, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 335, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 309, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 297, in process_inputs
result = f(**inputs)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\comfy_api\latest\_io.py", line 1833, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
```
## System Information
- ComfyUI Version: 0.22.3
- Arguments: G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\main.py --auto-launch --preview-method auto --use-sage-attention --disable-cuda-malloc
- OS: win32
- Python Version: 3.13.11 (tags/v3.13.11:6278944, Dec 5 2025, 16:26:58) [MSC v.1944 64 bit (AMD64)]
- Embedded Python: false
- PyTorch Version: 2.9.1+cu130
## Devices
- Name: cuda:0 NVIDIA GeForce RTX 5060 Ti : cudaMallocAsync
- Type: cuda
- VRAM Total: 17102864384
- VRAM Free: 15908995072
- Torch VRAM Total: 0
- Torch VRAM Free: 0
## Logs
```
2026-06-01T11:24:48.633244 - OSError: [WinError 121] 信号灯超时时间已到
2026-06-01T11:24:48.633351 -
The above exception was the direct cause of the following exception:
2026-06-01T11:24:48.633371 - Traceback (most recent call last):
2026-06-01T11:24:48.634268 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\custom_nodes\ComfyUI-Manager\glob\manager_server.py", line 1817, in get_cache
json_obj = await manager_util.get_data(uri, True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.634330 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\custom_nodes\ComfyUI-Manager\glob\manager_util.py", line 192, in get_data
async with session.get(uri, headers=headers) as resp:
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.634380 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 1510, in aenter
self._resp: RetType = await self.coro
^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.634428 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 779, in _request
resp = await handler(req)
^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.634475 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 757, in connectand_send_request
await resp.start(conn)
2026-06-01T11:24:48.634522 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client_reqrep.py", line 539, in start
message, payload = await protocol.read() # type: ignore[union-attr]
^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.634559 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\streams.py", line 703, in read
await self._waiter
2026-06-01T11:24:48.634590 - aiohttp.client_exceptions.ClientOSError: [WinError 121] 信号灯超时时间已到
2026-06-01T11:24:48.634696 - [ComfyUI-Manager] Failed to perform initial fetching 'github-stats.json': [WinError 121] 信号灯超时时间已到
2026-06-01T11:24:48.635255 - Traceback (most recent call last):
2026-06-01T11:24:48.635473 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\asyncio\proactor_events.py", line 286, in loopreading
length = fut.result()
2026-06-01T11:24:48.635524 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\asyncio\windows_events.py", line 804, in _poll
value = callback(transferred, key, ov)
2026-06-01T11:24:48.635573 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\asyncio\windows_events.py", line 463, in finish_socket_func
return ov.getresult()
~~~~~~~~~~~~^^
2026-06-01T11:24:48.635606 - OSError: [WinError 121] 信号灯超时时间已到
2026-06-01T11:24:48.635630 -
The above exception was the direct cause of the following exception:
2026-06-01T11:24:48.635648 - Traceback (most recent call last):
2026-06-01T11:24:48.635985 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\custom_nodes\ComfyUI-Manager\glob\manager_server.py", line 1817, in get_cache
json_obj = await manager_util.get_data(uri, True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.636038 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\custom_nodes\ComfyUI-Manager\glob\manager_util.py", line 192, in get_data
async with session.get(uri, headers=headers) as resp:
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.636086 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 1510, in aenter
self._resp: RetType = await self.coro
^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.636121 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 779, in _request
resp = await handler(req)
^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.636155 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client.py", line 757, in connectand_send_request
await resp.start(conn)
2026-06-01T11:24:48.636201 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\client_reqrep.py", line 539, in start
message, payload = await protocol.read() # type: ignore[union-attr]
^^^^^^^^^^^^^^^^^^^^^
2026-06-01T11:24:48.636236 - File "G:\ComfyUI-aki\ComfyUI-aki-v3\python\Lib\site-packages\aiohttp\streams.py", line 703, in read
await self._waiter
2026-06-01T11:24:48.636263 - aiohttp.client_exceptions.ClientOSError: [WinError 121] 信号灯超时时间已到
2026-06-01T11:24:57.581508 - FETCH ComfyRegistry Data: 5/1502026-06-01T11:24:57.581546 -
2026-06-01T11:25:06.489231 - FETCH ComfyRegistry Data: 10/1502026-06-01T11:25:06.489269 -
2026-06-01T11:25:15.464641 - FETCH ComfyRegistry Data: 15/1502026-06-01T11:25:15.464675 -
2026-06-01T11:25:24.980617 - FETCH ComfyRegistry Data: 20/1502026-06-01T11:25:24.980655 -
2026-06-01T11:25:35.217302 - FETCH ComfyRegistry Data: 25/1502026-06-01T11:25:35.217336 -
2026-06-01T11:25:45.411049 - FETCH ComfyRegistry Data: 30/1502026-06-01T11:25:45.411086 -
2026-06-01T11:25:54.324220 - FETCH ComfyRegistry Data: 35/1502026-06-01T11:25:54.324445 -
2026-06-01T11:26:03.474513 - FETCH ComfyRegistry Data: 40/1502026-06-01T11:26:03.474549 -
2026-06-01T11:26:12.769918 - FETCH ComfyRegistry Data: 45/1502026-06-01T11:26:12.769953 -
2026-06-01T11:26:21.544115 - FETCH ComfyRegistry Data: 50/1502026-06-01T11:26:21.544154 -
2026-06-01T11:26:30.904323 - FETCH ComfyRegistry Data: 55/1502026-06-01T11:26:30.904358 -
2026-06-01T11:26:40.855013 - FETCH ComfyRegistry Data: 60/1502026-06-01T11:26:40.855049 -
2026-06-01T11:26:50.427375 - FETCH ComfyRegistry Data: 65/1502026-06-01T11:26:50.427413 -
2026-06-01T11:27:00.189835 - FETCH ComfyRegistry Data: 70/1502026-06-01T11:27:00.189871 -
2026-06-01T11:27:09.203895 - FETCH ComfyRegistry Data: 75/1502026-06-01T11:27:09.203949 -
2026-06-01T11:27:18.989799 - FETCH ComfyRegistry Data: 80/1502026-06-01T11:27:18.989865 -
2026-06-01T11:27:28.090104 - FETCH ComfyRegistry Data: 85/1502026-06-01T11:27:28.090176 -
2026-06-01T11:27:36.896360 - FETCH ComfyRegistry Data: 90/1502026-06-01T11:27:36.896396 -
2026-06-01T11:27:45.853314 - FETCH ComfyRegistry Data: 95/1502026-06-01T11:27:45.853382 -
2026-06-01T11:27:55.474401 - FETCH ComfyRegistry Data: 100/1502026-06-01T11:27:55.474470 -
2026-06-01T11:28:04.814856 - FETCH ComfyRegistry Data: 105/1502026-06-01T11:28:04.814939 -
2026-06-01T11:28:14.329256 - FETCH ComfyRegistry Data: 110/1502026-06-01T11:28:14.329344 -
2026-06-01T11:28:23.406689 - FETCH ComfyRegistry Data: 115/1502026-06-01T11:28:23.406797 -
2026-06-01T11:28:33.271903 - FETCH ComfyRegistry Data: 120/1502026-06-01T11:28:33.272174 -
2026-06-01T11:28:42.593375 - FETCH ComfyRegistry Data: 125/1502026-06-01T11:28:42.593419 -
2026-06-01T11:28:52.461313 - FETCH ComfyRegistry Data: 130/1502026-06-01T11:28:52.461363 -
2026-06-01T11:29:01.875782 - FETCH ComfyRegistry Data: 135/1502026-06-01T11:29:01.875909 -
2026-06-01T11:29:11.741492 - FETCH ComfyRegistry Data: 140/1502026-06-01T11:29:11.741627 -
2026-06-01T11:29:22.200312 - FETCH ComfyRegistry Data: 145/1502026-06-01T11:29:22.200382 -
2026-06-01T11:29:31.124824 - FETCH ComfyRegistry Data: 150/1502026-06-01T11:29:31.124870 -
2026-06-01T11:29:31.625874 - FETCH ComfyRegistry Data [DONE]2026-06-01T11:29:31.625912 -
2026-06-01T11:29:31.787588 - [ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
2026-06-01T11:29:31.818562 - FETCH DATA from: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json2026-06-01T11:29:32.546369 - [DONE]2026-06-01T11:29:32.546414 -
2026-06-01T11:29:32.625487 - [ComfyUI-Manager] All startup tasks have been completed.
2026-06-01T12:08:33.639133 - got prompt
2026-06-01T12:08:33.660606 - Failed to validate prompt for output 2277:3278:
2026-06-01T12:08:33.660846 - * (prompt):
2026-06-01T12:08:33.660943 - - Required input is missing: source
2026-06-01T12:08:33.661061 - * PreviewAny 2277:3278:
2026-06-01T12:08:33.661145 - - Required input is missing: source
2026-06-01T12:08:33.661226 - Output will be ignored
2026-06-01T12:08:33.661376 - Failed to validate prompt for output 3185:
2026-06-01T12:08:33.661467 - * (prompt):
2026-06-01T12:08:33.661581 - - Required input is missing: images
2026-06-01T12:08:33.661662 - * SaveImage 3185:
2026-06-01T12:08:33.661695 - - Required input is missing: images
2026-06-01T12:08:33.661724 - Output will be ignored
2026-06-01T12:08:34.639991 - CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
2026-06-01T12:08:34.725792 - Found quantization metadata version 1
2026-06-01T12:08:34.725900 - Detected mixed precision quantization
2026-06-01T12:08:34.726099 - Using mixed precision operations
2026-06-01T12:08:34.726158 - Native ops: float8_e4m3fn, nvfp4, float8_e5m2 , emulated ops: mxfp8
2026-06-01T12:08:34.785826 - model weight dtype torch.bfloat16, manual cast: torch.bfloat16
2026-06-01T12:08:34.806268 - model_type FLUX
2026-06-01T12:08:36.465849 - VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
2026-06-01T12:08:37.257234 - VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
2026-06-01T12:08:37.305665 - Using sage attention mode: auto
2026-06-01T12:08:37.307487 - Applying LTX2 Memory Efficient Sage Attention Patch to all transformer blocks, triton_kernels=True
2026-06-01T12:08:37.311693 - Applying LTX2 Attention Tuner Patch with custom scales to blocks: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47], triton_kernels=True
2026-06-01T12:08:37.316101 - [rgthree-comfy][Power Lora Loader]2026-06-01T12:08:37.316148 - Matched Lora input "LTX/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors" to "ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors".2026-06-01T12:08:37.316169 -
2026-06-01T12:08:37.866716 - Requested to load LTXAVTEModel_
2026-06-01T12:08:38.043004 - Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
2026-06-01T12:08:52.778102 - 0 models unloaded.
2026-06-01T12:08:52.824041 - Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
2026-06-01T12:08:55.977149 - Requested to load VideoVAE
2026-06-01T12:08:56.116092 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:08:57.066529 - VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
2026-06-01T12:09:00.109086 - Requested to load LTXAV
2026-06-01T12:09:01.557915 - 0 models unloaded.
2026-06-01T12:09:01.812281 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:09:01.817554 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:12:10.773565 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:12:13.445821 - Requested to load VideoVAE
2026-06-01T12:12:13.482819 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:12:14.685546 - 0 models unloaded.
2026-06-01T12:12:14.839617 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:12:14.862389 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:16:06.567448 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:16:06.800620 - Requested to load AudioVAE
2026-06-01T12:16:09.095818 - loaded completely; 693.46 MB loaded, full load: True
2026-06-01T12:16:10.297513 - 0 models unloaded.
2026-06-01T12:16:10.306212 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:17:27.416405 - Prompt executed in 533.73 seconds
2026-06-01T12:31:41.775763 - got prompt
2026-06-01T12:31:41.819691 - Failed to validate prompt for output 2277:3278:
2026-06-01T12:31:41.819892 - * (prompt):
2026-06-01T12:31:41.819987 - - Required input is missing: source
2026-06-01T12:31:41.820106 - * PreviewAny 2277:3278:
2026-06-01T12:31:41.820191 - - Required input is missing: source
2026-06-01T12:31:41.820267 - Output will be ignored
2026-06-01T12:31:41.820382 - Failed to validate prompt for output 3185:
2026-06-01T12:31:41.820454 - * (prompt):
2026-06-01T12:31:41.820522 - - Required input is missing: images
2026-06-01T12:31:41.821078 - * SaveImage 3185:
2026-06-01T12:31:41.821454 - - Required input is missing: images
2026-06-01T12:31:41.821589 - Output will be ignored
2026-06-01T12:31:43.286783 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:31:44.981106 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:31:44.985333 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:32:54.710016 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:32:59.040047 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:32:59.658185 - 0 models unloaded.
2026-06-01T12:33:00.001511 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:33:00.006037 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:33:37.578178 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:33:37.641677 - Requested to load AudioVAE
2026-06-01T12:33:40.533030 - loaded completely; 693.46 MB loaded, full load: True
2026-06-01T12:33:41.664654 - 0 models unloaded.
2026-06-01T12:33:41.679170 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:34:08.033312 - Prompt executed in 146.18 seconds
2026-06-01T12:38:24.948783 - got prompt
2026-06-01T12:38:24.986018 - Failed to validate prompt for output 2277:3278:
2026-06-01T12:38:24.986249 - * (prompt):
2026-06-01T12:38:24.986399 - - Required input is missing: source
2026-06-01T12:38:24.986543 - * PreviewAny 2277:3278:
2026-06-01T12:38:24.986698 - - Required input is missing: source
2026-06-01T12:38:24.986816 - Output will be ignored
2026-06-01T12:38:24.986975 - Failed to validate prompt for output 3185:
2026-06-01T12:38:24.987068 - * (prompt):
2026-06-01T12:38:24.987157 - - Required input is missing: images
2026-06-01T12:38:24.987246 - * SaveImage 3185:
2026-06-01T12:38:24.987322 - - Required input is missing: images
2026-06-01T12:38:24.987394 - Output will be ignored
2026-06-01T12:38:25.482145 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:38:26.949981 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:38:26.953197 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:39:22.511067 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:39:24.842351 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:39:25.483049 - 0 models unloaded.
2026-06-01T12:39:25.792550 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T12:39:25.799654 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T12:40:10.815568 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T12:40:10.882727 - Requested to load AudioVAE
2026-06-01T12:40:11.474359 - loaded completely; 693.46 MB loaded, full load: True
2026-06-01T12:40:12.609179 - 0 models unloaded.
2026-06-01T12:40:12.624452 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T12:40:32.396634 - Prompt executed in 127.38 seconds
2026-06-01T13:08:53.235284 - got prompt
2026-06-01T13:08:53.288902 - Failed to validate prompt for output 2196:
2026-06-01T13:08:53.289123 - * ColorMatchV2 2277:3090:
2026-06-01T13:08:53.289238 - - Required input is missing: image_ref
2026-06-01T13:08:53.289344 - Output will be ignored
2026-06-01T13:08:53.289788 - Failed to validate prompt for output 2277:3278:
2026-06-01T13:08:53.289925 - * (prompt):
2026-06-01T13:08:53.290027 - - Required input is missing: source
2026-06-01T13:08:53.290200 - * PreviewAny 2277:3278:
2026-06-01T13:08:53.290302 - - Required input is missing: source
2026-06-01T13:08:53.290418 - Output will be ignored
2026-06-01T13:08:53.290555 - Failed to validate prompt for output 3185:
2026-06-01T13:08:53.290638 - * (prompt):
2026-06-01T13:08:53.290711 - - Required input is missing: images
2026-06-01T13:08:53.290802 - * SaveImage 3185:
2026-06-01T13:08:53.290876 - - Required input is missing: images
2026-06-01T13:08:53.290952 - Output will be ignored
2026-06-01T13:08:53.291177 - Failed to validate prompt for output 2182:
2026-06-01T13:08:53.291280 - Output will be ignored
2026-06-01T13:08:53.706921 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T13:08:54.682456 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T13:08:54.685850 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T13:09:01.636610 - Interrupting prompt 7efee175-0294-4151-8cc1-9f5d93bf8ae1
2026-06-01T13:09:01.641979 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T13:09:01.652156 - Processing interrupted
2026-06-01T13:09:01.654724 - Prompt executed in 8.33 seconds
2026-06-01T13:09:21.274255 - got prompt
2026-06-01T13:09:21.299423 - Failed to validate prompt for output 2196:
2026-06-01T13:09:21.299626 - * ColorMatchV2 2277:3090:
2026-06-01T13:09:21.299736 - - Required input is missing: image_ref
2026-06-01T13:09:21.299836 - Output will be ignored
2026-06-01T13:09:21.300178 - Failed to validate prompt for output 2277:3278:
2026-06-01T13:09:21.300288 - * (prompt):
2026-06-01T13:09:21.300378 - - Required input is missing: source
2026-06-01T13:09:21.300476 - * PreviewAny 2277:3278:
2026-06-01T13:09:21.300561 - - Required input is missing: source
2026-06-01T13:09:21.300641 - Output will be ignored
2026-06-01T13:09:21.300761 - Failed to validate prompt for output 3185:
2026-06-01T13:09:21.300850 - * (prompt):
2026-06-01T13:09:21.300923 - - Required input is missing: images
2026-06-01T13:09:21.301009 - * SaveImage 3185:
2026-06-01T13:09:21.301078 - - Required input is missing: images
2026-06-01T13:09:21.301152 - Output will be ignored
2026-06-01T13:09:21.301379 - Failed to validate prompt for output 2182:
2026-06-01T13:09:21.301478 - Output will be ignored
2026-06-01T13:09:25.423793 - Applying LTX2 Attention Tuner Patch with custom scales to blocks: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47], triton_kernels=True
2026-06-01T13:09:25.430695 - [rgthree-comfy][Power Lora Loader]2026-06-01T13:09:25.430792 - Matched Lora input "LTX/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors" to "ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors".2026-06-01T13:09:25.430841 -
2026-06-01T13:09:25.878337 - Requested to load LTXAVTEModel_
2026-06-01T13:09:26.020973 - 0 models unloaded.
2026-06-01T13:09:26.246383 - Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
2026-06-01T13:09:29.806677 - Interrupting prompt a530d13e-465b-427f-8abb-b710db70fedc
2026-06-01T13:09:29.845628 - Processing interrupted
2026-06-01T13:09:29.847460 - Prompt executed in 8.52 seconds
2026-06-01T13:09:37.837872 - got prompt
2026-06-01T13:09:37.867781 - Failed to validate prompt for output 2196:
2026-06-01T13:09:37.867976 - * ColorMatchV2 2277:3090:
2026-06-01T13:09:37.868077 - - Required input is missing: image_ref
2026-06-01T13:09:37.868170 - Output will be ignored
2026-06-01T13:09:37.868506 - Failed to validate prompt for output 2277:3278:
2026-06-01T13:09:37.868625 - * (prompt):
2026-06-01T13:09:37.868713 - - Required input is missing: source
2026-06-01T13:09:37.868806 - * PreviewAny 2277:3278:
2026-06-01T13:09:37.868888 - - Required input is missing: source
2026-06-01T13:09:37.868962 - Output will be ignored
2026-06-01T13:09:37.869078 - Failed to validate prompt for output 3185:
2026-06-01T13:09:37.869166 - * (prompt):
2026-06-01T13:09:37.869248 - - Required input is missing: images
2026-06-01T13:09:37.869340 - * SaveImage 3185:
2026-06-01T13:09:37.869424 - - Required input is missing: images
2026-06-01T13:09:37.869491 - Output will be ignored
2026-06-01T13:09:37.869684 - Failed to validate prompt for output 2182:
2026-06-01T13:09:37.869766 - Output will be ignored
2026-06-01T13:09:38.460406 - Applying LTX2 Attention Tuner Patch with custom scales to blocks: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47], triton_kernels=True
2026-06-01T13:09:38.464710 - [rgthree-comfy][Power Lora Loader]2026-06-01T13:09:38.464789 - Matched Lora input "LTX/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors" to "ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors".2026-06-01T13:09:38.464845 -
2026-06-01T13:09:38.795516 - Requested to load LTXAVTEModel_
2026-06-01T13:09:38.924605 - 0 models unloaded.
2026-06-01T13:09:39.008631 - Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
2026-06-01T13:09:55.559640 - 0 models unloaded.
2026-06-01T13:09:55.639894 - Model LTXAVTEModel_ prepared for dynamic VRAM loading. 25440MB Staged. 0 patches attached. Force pre-loaded 290 weights: 1497 KB.
2026-06-01T13:09:59.768402 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T13:10:00.149450 - Requested to load LTXAV
2026-06-01T13:10:01.617448 - 0 models unloaded.
2026-06-01T13:10:02.040394 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T13:10:02.045151 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T13:11:55.916510 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T13:12:00.312398 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T13:12:01.033058 - 0 models unloaded.
2026-06-01T13:12:01.356457 - Model LTXAV prepared for dynamic VRAM loading. 16778MB Staged. 1660 patches attached.
2026-06-01T13:12:01.364988 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
2026-06-01T13:12:49.162119 - Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
2026-06-01T13:12:49.236820 - Requested to load AudioVAE
2026-06-01T13:12:51.693173 - loaded completely; 693.46 MB loaded, full load: True
2026-06-01T13:12:52.898850 - 0 models unloaded.
2026-06-01T13:12:52.914379 - Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
2026-06-01T13:13:08.466807 - !!! Exception during processing !!! ColorMatchV2.execute() missing 1 required positional argument: 'image_ref'
2026-06-01T13:13:08.481369 - Traceback (most recent call last):
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 535, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 335, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 309, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\execution.py", line 297, in process_inputs
result = f(**inputs)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "G:\ComfyUI-aki\ComfyUI-aki-v3\ComfyUI\comfy_api\latest\_io.py", line 1833, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
TypeError: ColorMatchV2.execute() missing 1 required positional argument: 'image_ref'
2026-06-01T13:13:08.510105 - Prompt executed in 210.61 seconds
```
## Attached Workflow
Please make sure that workflow does not contain any sensitive information such as API keys or passwords.
```
Workflow too large. Please manually upload the workflow from local file system.
```
## Additional Context
(Please add any additional context or steps to reproduce the error here)
The workflow does not contain that node, it was replaced by color transfer, please update the workflow
@Darksidewalker Okay, thank you.
No conditions are connected to the LTX2 NAG node. Is this working? When I tried connecting a condition from Stage1 to it, an error occurred because the processing dependencies could not be resolved.
this workflow would be awesome with LTX director!!! thanks somcuh for your work im loving it!!
Thanks for this great Workflow, together with your "Golden Lace" model—simply fantastic!
first ive updated all nodes and comfyui to start and i disabled the ffmpeg video upload cause that just breaks everything. the second LTXVImgToVideoInplace is red and is skipping it it goes through ok but a video is not created with the video combine just laggy frames created in the submenu here are the errors [ERROR] Failed to validate prompt for output 2277:3289:
[ERROR] * (prompt):
[ERROR] - Required input is missing: source
[ERROR] * PreviewAny 2277:3289:
[ERROR] - Required input is missing: source
[ERROR] Output will be ignored
[ERROR] Failed to validate prompt for output 2196:
[ERROR] * LTXVImgToVideoInplace 2277:3205:
[ERROR] - Required input is missing: bypass
[ERROR] Output will be ignored
[ERROR] Failed to validate prompt for output 3185:
[ERROR] * (prompt):
[ERROR] - Required input is missing: images
[ERROR] * SaveImage 3185:
[ERROR] - Required input is missing: images
[ERROR] Output will be ignored
[ERROR] Failed to validate prompt for output 2182:
[ERROR] Output will be ignored
How are the [0.715, 0.4824, 0.2412, 0.0] values in the Sigmas Upscale node calculated? On my end, these numbers cannot generate a video; what is produced is a gray, jittery image.
There are now two questions:
1. Do these original sigma values depend on a specific runtime environment (such as the PyTorch version or GPU model)?
2. In two-stage sampling, is there a recommended number of steps and method for generating sigma values in the second stage (automatic scheduler vs fixed values)?
That's not a problem of these sigmas. You may have messed settings or confusing non-distilled settings vs distilled
@Darksidewalker Sorry, I just started using LTX2.3, and I'm not very familiar with the functions of many nodes and parameter settings. It took me three days of testing to identify that the problem was with the Sigmas Upscale node in the second sampling. I found two ways to solve this problem.
1. I copied an LTXVScheduler node from the first sampling to replace the Sigmas Upscale node, and the video could be generated normally, although it seems that the advantage of second sampling was not fully utilized.
2. Changing the values in the Sigmas Upscale node to [0.987, 0.85, 0.725, 0.422, 0.0] also allowed the video to be generated normally, but the quality of the generated video seems to be no different from the video generated by my other single-sampling workflow. This was the method suggested to me by DeepSeek.
DeepSeek also mentioned a possibility: because of different versions of ComfyUI and some plugins, there may be issues with values being changed when loading a workflow. For example, if a node’s value was originally set to 0.2, due to version differences, it could be loaded as 2.
@1191300911 I have no clue what to tell you, but deepseek is wrong and ether your comfyui installation is broken or you did wrong settings. You do not have to change the nodes in the backend.
i am trying to test the workflow and disabling the audio, but it seems putting placeholder.mp3 and placeholder.mp4 only bypasses the validation and stuck at 67%
Hi! Your workflow is amazing. The only change I want to make is to load separate loras for the model audio. However, I'm a little intimidated of the spaghetti subgraph and I don't want to screw anything up. Can some one kindly point me to where I can insert an additional power lora node and exactly what nodes it should connect to? Thanks!
the lora node can already load audio only.
you can mix any lora and just load the part you want.
Hmm, I'm not seeing it. Let me phrase it another way. My main issue is with the singularity-ltx omni lora. When it is implemented into the workflow I can see an improvement in visual fidelity but when the lora is active there is a noticeable degradation of the audio quality both in foley and voices. I want to be able to load the lora for the video generation but ignore it during the audio generation. Is this possible?
@Mastermike8800 In my recent v3.8 the dasiwa-lora-loader can disable the audio part entirely
@Darksidewalker thanks I'll check it out
@Darksidewalker Exactly what I need and worked like a charm.
@Darksidewalker i still dont think its compatible
do you have a updated version?
@DekotaDEE69 what do you mean, I did multiple videos with it, working fine.
@Darksidewalker the lora loader itself, i cannot load loras? there is no way to add lora
@DekotaDEE69 click the name field or + button
@Darksidewalker itsok i just downlaoded again and its working fine! dont know what happened. thanks os much!! what a great workflow!! i love it!!!!
Wow! Cool workflow upgrade, thanks!
the Dasiwa LTX lora loader isnt working for me. it seems thats its not compatible
ComfyUI 0.21.0ComfyUI_frontend v1.43.18Templates v0.9.73DiscordComfyOrgEasyUse v1.3.6rgthree-comfy v1.0.2605082257
System Info
Copy System Info
OS
win32
Python Version
3.13.12 (tags/v3.13.12:1cbe481, Feb 3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]
Embedded Python
true
Pytorch Version
2.11.0+cu130
Arguments
ComfyUI\main.py --windows-standalone-build --enable-manager
RAM Total
61.63 GB
RAM Free
42.17 GB
Templates Version
0.9.73
the director is awesojme addition, im using a differnt lora loader ofr now, would be awesome to be able to load IC lora , so we can use voice samples aswell! this workflow is pretty much the best out there! this addition would make it unbeatable!!!
Do you archive your older versions anywhere? The director is an amazing addition to the LTX workflow, but as you know it comes with the loss of V2V. It would be nice to have an old V2V version of the workflow in pocket should the need arise
Near the bottom of the description there's a Github link with all of them archived
Anyone else having trouble getting the rtx settings/nodes to work properly? getting this error
[ERROR] !!! Exception during processing !!! Failed to create NVIDIA RTX VFX effect (Denoise DENOISE_ULTRA): NvVFX_Load failed: The effect has not been properly initialized (code -12) [ERROR] Traceback (most recent call last): File "I:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-DaSiWa-Nodes\nodes\rtx_upscaler_refiner.py", line 283, in _maybe_vfx_effect effect.load()Been troubleshooting for a few hours so curious if this is a me issue or if anyone else is experiencing this? Thanks!
Make sure you installed the nodepack and requirements correctly.
For anyone experiencing this same issue in the future, rolling back from nvidia driver version 610.47 back to 596.49 fixed it
Since the latest ComfyUI update, I'm having an issue with the workflows. It seems like something is missing. The part where the resolution is set just says 'disconnected'—this happens with the current as well as with older workflows. I have already updated all custom nodes and even did a fresh install of ComfyUI, but the issue persists.
I had the same “Disconnected” issue with the DaSiWa workflow after the recent ComfyUI update.
This workaround fixed it for me:
1. First select the affected node, for example the “Settings and Backend” node.
2. Click the “Toggle properties panel” button to open the Properties panel/sidebar.
3. In the opened Properties panel, use the settings/filter button in the top right of that panel.
4. Search for the missing settings, for example resolution_preset, or other settings that disappeared from the node.
5. Click the eye icon next to the missing setting to show it on the node again.
6. The restored setting may appear at the bottom of the node first. You may need to resize the node and/or reorder the setting in the Properties panel.
7. After doing this, the “Disconnected” fields disappeared for me and the missing settings were visible again.
Funny enough, I only found this by accident after I misunderstood/misphrased something while troubleshooting 😂
Hope this helps!
same but i still cant get it to work
I really appreciate you incorporating the LTX Director node into your workflows. It takes a little longer to set up, but less renders to get it right! Thanks!!
Can we have a 2 different options in this workflow: LTX Director and the usual approach? Director is cool and so on, but it's hard to connect this Director thing to my other complex pipelines, and it's kinda overwhelming, I would like to stick to old fashioned way for now. Thank you
Well, I'm doing the workflows in 1st place how I want to use them and just make them available for everyone, so I won't maintain multiple versions of it. Sorry.
I already do so much project, I cannot make everyone happy and fulfill all needs, that's just not possible for 1-person doing so much projects. You can use an older workflow, or change aspects you need, but I will stay away from maintaining multiple versions of everything.
If I feel the LTX director does not suite my needs, I'll change it, but for now it is like I use LTX23.
@Darksidewalker Got it!
Have you found a way to use this awesome workflow, but remove the LTX director?
awesome workflow friend, guys could you tell what i need to move for get better performance in the final video, i am getting a little bad quality, for example i get a very good quality with the ltx23AllInOneWorkflowForRTX_v20 workflow
Raise any resolution settling
@Darksidewalker Are you referring to the node DaSiWa Resolution Scale Calculator options or resolution preset??
Thanks a bunch, good stuff!
Oh you are my god!!! this workflows work very fast in my RTX 5060. thank you so much.
The FLF2V switch is not visible. How can I enable or use FLF2V?
Hey dude.
your workflow is really so great and easy to use..I can't go back to other WF hahaha
I would like to ask you, if you have time... Could you implement the new feature about that Licon-MSR reference ?
It seems to be something that enhance greatly consistency of characters for example.
Thank you for mentioning it, looks impressive. When I get time I'll have a look into it 😸
@DarksidewalkerThanks so much !!Looking forward to it !
Missing Node Packs (1)
This workflow uses custom nodes you haven't installed yet.
To install missing nodes, first run pip install -U --pre comfyui-manager in your Python environment to install Node Manager, then restart ComfyUI with the --enable-manager flag.
ComfyUI-LTXVideo (1)
LTXVSpatioTemporalTiledVAEDecode
Thank you for this workflow! It is the best workflow I have ever used for video. LTX Director is a game changer. It took awhile to get everything working, but worth it. For people having issues, I ended up going with a slightly older version of ComfyUI Desktop 0.9.4. That worked for me + installing all of the required nodes.
With the latest version of WhatDreamsCost-ComfyUI, a bunch of LTX director bugs pop up. Considering that I reinstalled ComfyUI with your installer.
[INFO] got prompt
[ERROR] Failed to validate prompt for output 2182:
[ERROR] * LTXDirector 3314:
[ERROR] - Value 0 smaller than min of 1: divisible_by
[ERROR] - Failed to convert an input value to a FLOAT value: frame_rate, seconds, could not convert string to float: 'seconds'
[ERROR] - Value not in list: display_mode: 900 not in ['frames', 'seconds']
[ERROR] * LTXVConditioning 3578:3564:
[ERROR] - Return type mismatch between linked nodes: frame_rate, received_type(MOTION_GUIDE_DATA) mismatch input_type(FLOAT)
[ERROR] * LTXDirectorGuide 3578:3560:
[ERROR] - Failed to convert an input value to a FLOAT value: ic_lora_strength, nearest-exact, could not convert string to float: 'nearest-exact'
[ERROR] - Value not in list: ic_lora_name: 0.5 not in (list of length 652)
[ERROR] * LTXDirectorGuide 3579:3552:
[ERROR] - Failed to convert an input value to a FLOAT value: ic_lora_strength, nearest-exact, could not convert string to float: 'nearest-exact'
[ERROR] - Value not in list: ic_lora_name: 1 not in (list of length 652)
[ERROR] * VHS_VideoCombine 2196:
[ERROR] - Return type mismatch between linked nodes: frame_rate, received_type(MOTION_GUIDE_DATA) mismatch input_type(FLOAT)
[ERROR] Output will be ignored
[ERROR] Failed to validate prompt for output 2196:
[ERROR] * (prompt):
[ERROR] - Return type mismatch between linked nodes: frame_rate, received_type(MOTION_GUIDE_DATA) mismatch input_type(FLOAT)
[ERROR] Output will be ignored
[ERROR] Failed to validate prompt for output 3608:
[ERROR] * (prompt):
[ERROR] - Required input is missing: images
[ERROR] * SaveImage 3608:
[ERROR] - Required input is missing: images
[ERROR] Output will be ignored
[WARNING] invalid prompt: {'type': 'prompt_outputs_failed_validation', 'message': 'Prompt outputs failed validation', 'details': 'Return type mismatch between linked nodes: frame_rate, received_type(MOTION_GUIDE_DATA) mismatch input_type(FLOAT)\nRequired input is missing: images', 'extra_info': {}}
Replace the LTX Director node and both LTX Director Guide nodes (inside the "1st Pass"/"2nd Pass" subgraphs — double-click to access). Then disconnect the wire feeding frame_rate on the LTXVConditioning node (it's wrongly wired to motion_guide_data) and set frame_rate manually instead. Hope this helps
Yes, I'm really being dumb. In addition, I also had to set ic_lora_name to the desired (or none) value in the LTX Director Guide, ic_lora_strenght to the value in both Passes, then it worked.
@himinglaeva47285 Can you explains more about what needed to be replaced when you say - "Replace the LTX Director node and both LTX Director Guide nodes (inside the "1st Pass"/"2nd Pass" subgraphs — double-click to access)."? I' don't know what I'm suppose to be replacing.
@nightime This is not relevant in the new versions of the workflow. There was a problem with upgrading to the new LTX director 2.0 version
On the video combine you need to set the fps to 24
I don't know why, but the videos are noticeably more blurred and pixelated than before on the same resolution settings.
Also, the 3rd pass takes 7-9 times longer than the 2nd one
Not sure why, but any time I generate a video w/ v 41 that has speech or any kind of dialogue in it, it gets completely bypassed/ignored. Other actions seem to work correctly though. Oddly enough, I tried the same exact prompts with your version 40 that's on your github and that one works perfectly. Maybe it's a problem with my machine?
There is no sound at all, and i have updated everything.
The default option for sound seem to be off, just enable. I may change that for next version
@vicaut Seems like audio inpaint has to be toggled on, even if you have no custom input.
My starting image is always 90 degrees flipped, regardless of the swap_aspect switch.
The input image was too big, i scaled it down and now it works correctly.
New to LTX workflow, so please patient if these are noob questions. I'm trying to run this workflow using the latest version of ComfyUI. I'm having these errors that I don't understand, any assistance is helpful!
RuntimeError: NVIDIA RTX VFX (nvvfx) module not found. Please ensure NVIDIA RTX Video SDK / Broadcast SDK is installed and the 'nvvfx' package is in your python path.
I did a manual in the embedded folder.
- python.exe -m pip install nvidia-vfx
and
-python.exe -m pip install wheel-stub
I also download the Nvidia_RTX_Nodes_ComfyUI through the manager. But nothing works. I still get the error above.
You also need to enable RTX Upscaling on the NVidia App. If not enabled there it will never work.
@Ai_Unleashed I tried the suggestion but it still doesn't work, same error.
I think the problem is that with the newest update, the name was changed to "nvidia-vfx" rather than "nvvfx".
I even tried to download the official node through manager, but it fails during installation because it can't find "nvvfx"
I quickly used Google: Found this. Looks like it could be interesting for you. https://www.youtube.com/watch?v=qmJbG0Hfiwo
They are talking about the same or at least similar Error...and how to fix it.
Not what I had to do but....
@Ai_Unleashed That's exactly what I did when I say I went manual in the embedded folder and installed nvidia-vfx. If I run that same code, it will said all requirements is there already. You see how his RTX node in the manager is blue, indicating it's installed and working. Mine is red saying that it can't find nvvfx. I really think it's the nvidia driver that is messing this up. I'm using the latest one. I saw in the comment here somewhere that the person reverted back to 5.9 to get it to work.
Great workflow, but I can't understand how to enable chunkling.
There is a note about Chunkling Headroom, but there is NO such setting. I opened every single subgraph and even used search in it with no result. In chuckling subgraph there is a note about VRAM Headroom, but again, I don't see what is it talking about since there is no a,b,c,d options there and nowhere to set those values.
I tried to toggle chunckling manually, to change chunks value, but it never splits anything and stucks at long videos.
What am I doing wrong? Is there something I don't see?
Chunking is for VRAM reduction not splitting video into longer chunks.
Anyone having issues loading LTX Director?
I had to re-creates the LTX Director node + LTX Director Guide nodes (in subgraph) to get to work.
Clearing browser cache and redownload the wf, sometimes cache is using old node settings
bookmark 2 ( I assume backend) is missing...I can run the WF and it generates a vid of a woman...but I cant navigate to bookmark 2...is this a me problem or does anybody else have this/know a solution?
its great Workflow but i had problem with [Errno 22] on VHS node (video COmbine)
BEST CREATOR EVER!!!!!
Ok, I've been using these OmniForge workflows successfully for a few months now, however, I have always bypassed the 'color transfer' option, I cannot for the life of me find that node in Comfyui. I have the Comfyui Post Processing nodes suite installed, but your workflows always shows this node as a red X and it does not appear when searched. What the heck am I missing here? I tried replacing with the colormatch node, but that breaks the workflow if enabled. Does anyone understand what the heck is going on with this node?
The color transfer is a native node since comfyui 0.21.
@cstudevan100 How should I know, it is a comfyui native node. So your installation must be faulty. 🤷
@Darksidewalker looks like i'm on version 0.19, so that may be why
@Darksidewalker Problem solved, had to do a git --reset hard command to get it to clear and update properly, thanks for the heads-up to point me to the problem
Could you please add the 2x FPS option back? It's really useful for highly dynamic scenes. I tried setting it directly to 48 FPS, but the output video had a stuttering effect.
I removed it, because it made issues with director. I can try again if I get time.
Node 'ID #3676:3631' has no class_type. The workflow may be corrupted or a custom node is missing.
Install latest comfyui and all custom nodes.
@Darksidewalker thx for the quick reply, there was some deeper bug all my stuff was up to date, claude code fixed what ever it was for me. im not sure what has changed but your older workflow was super quick like 200s for a 5 sec video, this one is taking forever. 5080 and 64gig of ram
@fellowricky 3 sampler stages, disable what you don't need, but it will reduce quality
Another suggestion: https://github.com/TenStrip/10S-Comfy-nodes
Helps a lot for face consistency.
In version 4.2 note has a link to gemma-3-12b-it-heretic-v2_fp8_e4m3fn.safetensors but the node uses just gemma_3_12B_it_fp8_e4m3fn.safetensors without it-heretic-v2 part in the name.
Are they different? Which one should we use? Which one is better? Thank you
All work. I did not compare, so I cannot tell the difference.
Whatever optimizations you did in 4.2 are great: even with three passes its a shorter generation time than the 3.x running two.
Thank you :)
The quality with the new 4.2 update is really good, you did a great job! ❤
I was wondering though if there’s any way to reduce rendering time a bit. I have an RTX 3090, and generating 5-second clips at 1280x720 takes me around 290 seconds. I got there after a some testing
For now, I’ve disabled the 3rd pass and I set Chunking to 1. Is there anything else I can do to reduce the render time?
First, thank you for the amazing workflow!
This is probably me doing something really dumb, but I cannot for the life of me get the audio inpainting to work correctly. If I put an existing generated video in and want to continue it, the 'inpainted' items before or after the video don't match the same audio as the video being extended. Audio IS generated (so inpainting is 'on'), but it's a suddenly jarringly different voice / different background noise, etc. I thought this might be an issue with LTX Director but when using the LTX Director workflows directly as provided by WhatDreamsCost, I don't run into that issue, so it has something to do with the way this workflow (which is otherwise amazing!) incorporates LTX Director.
Any idea where to start for troubleshooting what's causing this??
the sage attention toggle for the 4.2 workflow throws a giant error
Not on my side, worked as intended. Make sure to install sage-attention properly.
@Darksidewalker yea its installed correctly
@PastelPastelPastel Than there would not be an error 🤷
Hi thanks for your flow as always :3 Just wondering there why there are no more extract last frame option ? Do you plan to reintroduce it later ?
The workflow fails because the model input of the LoRA Loader node is missing a required connection.
It does not. I use it every day, make sure you did not install broken custom nodes that delete connections like translate nodes.
same issue
@guuucat same answer
Simply the best! Nice and clean; easy to understand and highly functional. This dude knows how to make a good workflow. :)
I'm getting an error in the subgraph -> Sample Custom Advanced Node. Anyone else getting the following?
EDIT: I tried the older V41 workflow from GITHUB and didn't get the same error.
## Error Details - Node ID: 3676:3650 - Node Type: SamplerCustomAdvanced - Exception Type: AttributeError - Exception Message: AttributeError: 'AutoencodingEngine' object has no attribute 'per_channel_statistics'
RuntimeError: Input type (float) and bias type (struct c10::BFloat16) should be the same.What is the problem?
Is there a working WF without Director please? I had an old version but is stopped working.
Maybe some old from the git repo
I just figured out that if you right click the image import in Director V2, you can select if it is a "start" frame or an "end" frame - this lets you create smooth looping videos:
https://civitai.red/posts/29781801
might be common knowledge, but I didn't know so posting it here just in case!
Currently im trying out ltx 2.3 more.u are using euler ancestral pp and linear quadratic.
(Rtx5090)If u are using lcm and sgm uniform the render time is dropping 20-50%.eyes and so on are morphing more that way.google gemini said there is a special node for that TCD noder or something.i have to look into that.i will try it myself these days.
But lcm is definitely way way faster(im using 8 step 1 phase)
hmm with the new workflow there is no improvments to switch to lcm/sgm uniform.
who knows maybe something bugged out on my "old" workflow....
I've recently updated my nodes for comfy, and it seems like it broke a part of the workflow, specifically all of the settings for the resolution/aspect ration scaling. They all appear as "Disconnected" instead of showing the settings. Should I do something in the backend to fix it?
You can technically still use them, just go into the subgraph and Bypass or Unbypass the ones you want to use. A hassle, but workable. I've found that some of the WF of DaSiWa for Wan and LTX the older ones are a lot faster, and smoother still.
Or you update comfy + all custom nodes, redownload the workflow and clear browser cahce.
Make sure not to use broken custom nodes like some translation nodes.
@Darksidewalker turns out I was using an outdated version of the workflow, so I downloaded the latest one. New problem: I'm trying to do img2vid, but the end result is just becoming a blob. Is color reference supposed to be the first frame equiavalent, because I have it enabled, and uploaded the picture, but in the video the first frame doesn't even show up.
@Darksidewalker Ok, managed to fix it, for some reason the director node was bugged and it wasn't displaying anything except the positive prompt. Just had to restart my PC.
Fantastic and easy to use workflow. I use it with LTX 2.3 fp8 with zero issues.
I kept getting blurry result. I mean VERY VERY blurry result. The audio is completely fine though.
Why is it? Is there option I can toggle to improve the result?
Same. I stay here for a help :>
I had the same issue. My results were extremely blurry at first, but it was fixed by turning on Distillation/Lightspeed in the Model Loader. It is disabled by default.
Make sure the DMD/distilled LoRA is selected, set the strength to 1.0, and enable Distillation/Lightspeed. That fixed the blur for me.
I was also using CFG 1 and euler_cfg_pp.
Starting off I want to say that you are a genius and I LOVE your workflows. Most of the time they are extremely optimized. I did find something other users might find handy to improve these workflows in particular. If you have the previous version of DaSiWa LTX OmniForge C-LTX23-12.json (which is my favorite by far), for the V2V you have to rewire the Separate Latent THRU the Crop Guides and THEN into spatial. Crispy results. Reminder that however many seconds you put in duration, is what will be extended on the uploaded video.
For this new workflow (and also for the old one) I place a Get Image from Batch node as a second link from the Upscaler (batch index -1 / length 1) to get the very last frame of the video, since the old version had a toggle but it extracted it from the already created video which increased the deterioration exponentially. It will still deteriorate with this method, but it's much better. Sidenote: Turn off Comfy Nodes 2.0 if your workflows went bananas.
Will you implement https://github.com/TenStrip/10S-Comfy-nodes#ltx-face-identity-reinforcer? Then it would be possible to use https://huggingface.co/Alissonerdx/LTX-Best-Face-ID for I2V.
Chances are low since incompatible with LTX-Director multi frames atm
I'm having troubles using it without a distiller, my video keeps coming out under baked and all messed up, seems like its not doing enough steps. All I did was turn the "distillation" switch off. Can someone tell me what I'm doing wrong? Did i miss an additional step?
If you use a model that needs distillation this is normal
AMAZING!!!!!!!!!!!!
My video always distorts after 5 seconds. Is this normal? If not, how to fix it?
The new pixel feature for inappropriate content 😁Maybe you need distilled checkpoint/Lora or appropriate settings
@Darksidewalker I am using DaSiWa-LTX23-GoldenLace-v3_fp8.safetensor. Should I use the distilled lora or not? Because I've been getting this distorted videos whether its SFW or NSFW.
@theaibakaboy v3 needs a distilled lora
I tested with and without DMD distill Lora with the v4 dragonleap model and this wf and I'm not sure it does anything. What is your best practice ? Use it ?
I resolved my issues by using the ltx-2.3-22b-distilled lora-1.1_fro90_ceil72_condsafe.safetensors at 0.7 instead of the DMD and it works much better. Also swapped sampler for euler cfg pp for better quality (but it's slower)
For some reason the Triton VAE distorts the audio for me (voices distorted, ambient sounds are just horrible noise). used Dragonleap and ErosV12 models. Audio is normal in Tiled VAE though.
Edit: nevermind, tiled VAE does it too on some videos. I have no idea now why it's happening 😩
I don't know what's going on... I was able to use v3.9 of this workflow no problems. Then I updated to v4.6 (and updated all the custom nodes accordingly), but it's taking twice as long to do the first pass then I hit an OOM error
How do you guys get rid of the nsfw elements, I got uncensored video with no prompt at all at first run, with prompt added I get characters with weird gestures and jiggling private parts.
If you do not want NSFW, not using a NSFW enhanced checkpoint would work
@Darksidewalker I am using dragonleap v4 for its enhanced features and anime capabilities. I just want to generate sfw anime videos like the coin flipping girl and other video examples on this page's gallery, I have already set cfg value to 1.5 and added negative prompt, but Istill get weird nsfw gestures and big jiggling private areas, should I get the unmodified LTX 3.3 instead? Btw, with your amazing workflow, I can get 720p videos generated in under 10 minutes with RTX 3060 and 32GB system RAM using dragonleap Q8 gguf , so thank you for making this available.
@turtlemadmax448 you should drop gguf and use int8 ConvRot. NAG should work better than cfg raise. The checkpoint is really made for NSFW, but I did SFW stuff. But also ltx23 hallucinating a lot more than WAN.
@Darksidewalker Thanks a lot for your great help, i will check int8ConvRot. Have a nice one.