🚀 MrXin LTX 2.3 I2V EROS V6
This workflow has continuously evolved based on user feedback to deliver one of the strongest LTX 2.3 I2V experiences available. Thx everyone for the feedbacks.
V6 is a major leap forward, focusing on stronger motion preservation, better prompt adherence, and more user control while keeping excellent stability and VRAM efficiency.
🔥 What’s New in V6 (Compared to V5)
Added LTX2 NAG (Negative Attention Guidance) on both First Pass and Final Pass for significantly improved prompt following and motion quality.
Switched to VAELoader KJ nodes for Video VAE and Audio VAE (better compatibility and stability).
Improved Final Pass by repositioning the LTXVCropGuides node for cleaner latent processing.
Added flexible sigma system: Manual Sigmas (8 steps) or automatically generated sigmas via LTXVScheduler + Sigmas Sigmoid (12 steps).
Final Pass is Manual Sigmas only for stability.
Audio volume default set to 0 (neutral) with easy slider control.
Enhanced image conditioning using LTXVImgToVideoInplaceKJ in both passes.
Cleaner workflow layout, better group organization, and more helpful in-workflow tips.
✅ Core Features
Extremely strong, natural and consistent motion from First to Final Pass
Excellent anatomy, lighting, physics and NSFW detail
High-quality 24 FPS video with perfectly synced audio
Easy model switching between 10Eros V1 FP8 and Distilled 22B
Built-in Video Editor (RTX Super Resolution + nmkdSiaxCX upscaler + RIFE 48/60 FPS)
Optimized for 16GB VRAM setups and 32GB RAM
📥 Required Models + Direct Download Links
10Eros V1 FP8 (Main Checkpoint) → ltx2310eros_v1_FP8.safetensors
Distilled Model → ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
Text Encoder → gemma_3_12B_it_fp8_e4m3fn.safetensors
Text Projection → ltx-2.3_text_projection_bf16.safetensors
Video VAE → LTX23_video_vae_bf16.safetensors
Audio VAE → LTX23_audio_vae_bf16.safetensors
Preview VAE → taeltx2_3.safetensors
Spatial Upscaler → ltx-2.3-spatial-upscaler-x2-1.0.safetensors
Model Upscaler → nmkdSiaxCX_200k.safetensors
Distilled LoRAs (First & Second Pass)
📁 Folder Structure
ComfyUI/
├───📂 models/
│ ├───📂 diffusion_models/
│ │ └─── ltx2310eros_beta.safetensors
│ │ └─── ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
│ │
│ ├───📂 text_encoders/
│ │ └─── gemma_3_12B_it_fp8_e4m3fn.safetensors
│ │
│ ├───📂 clip/
│ │ └─── ltx-2.3_text_projection_bf16.safetensors
│ │
│ ├───📂 VAE/
│ │ └─── LTX23_video_vae_bf16.safetensors
│ │ └─── LTX23_audio_vae_bf16.safetensors
│ │ └─── taeltx2_3.safetensors
│ │
│ ├───📂 latent_upscale_models/
│ │ └─── ltx-2.3-spatial-upscaler-x2-1.1.safetensors
│ │
│ ├───📂 upscale_models/
│ │ └─── nmkdSiaxCX_200k.safetensors
│ │
│ └───📂 Lora/
│ └─── ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors
│ └─── ltx-2.3-22b-distilled-lora-384-1.1.safetensors
Custom Nodes Overview – MrXin LTX 2.3 I2V EROS V6
ComfyUI-KJNodes
rgthree-comfy
ComfyUI-easy-use
ComfyUI-mxToolkit
ComfyUI-VideoHelperSuite
ComfyUI-LTXVideo
controlaltai-nodes
comfyui_nvidia_rtx_nodes
GACLove/ComfyUI-VFI (voor RIFE)
comfyui_memory_cleanup + comfyui-impact-pack
Pro Tip: Use the new 10Eros V1 FP8 model with its recommended trigger words for the best results.
✅ Quick OOM Fix Guide
If you’re getting Out of Memory errors or weird faces in the final pass, follow these steps in order:
1. Update or Fresh Install ComfyUI + Install Missing Custom Nodes
Download and run the Easy Installer: 👉 https://github.com/Tavris1/ComfyUI-Easy-Install
Or watch this short video: 👉 https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
After updating/installing:
Open ComfyUI → go to Manager → click Install Missing Custom Nodes
Restart ComfyUI completely.
2. Add the Low VRAM Flags
Go to your ComfyUI folder.
Find the file ComfyUI.bat.
Right-click → Edit (open with Notepad).
Find the line that has ComfyUI\main.py .
Add the flags at the end so it looks like this:
bat
ComfyUI\main.py --lowvram --reserve-vram 6 --preview-method none --disable-xformers --disable-smart-memorySave and launch ComfyUI using this .bat file.
3. Increase Windows Swap File (Virtual Memory) This helps when VRAM still spikes:
Right-click This PC → Properties → Advanced system settings.
Click Settings under Performance → Advanced tab → Change (under Virtual memory).
Uncheck “Automatically manage paging file size for all drives”.
Select your main drive (usually C:) → choose Custom size.
Set:
Initial size (MB): 49152
Maximum size (MB): 98304
Click Set → OK → restart your PC.
After doing all three steps, reload the workflow and test with Longer Side 1024 and Video Length 20 seconds first. This combination fixes OOM for almost everyone on 12GB cards.
Version History & Previous Updates
V5 — Major base model upgrade to 10Eros V1 FP8. Significantly better realism, anatomy, lighting and motion coherence. Updated model loading and separate LoRA handling for each pass.
V4 — Introduced separate Manual Sigmas for First and Final Pass, much stronger Final Pass motion, full TwoWaySwitch system for model selection, expanded LoRA stack, and refined Video Editor.
V3 — Solved the biggest V2 complaint by preserving much more motion in the Final Pass. Added Model Upscaler (nmkdSiaxCX), option to disable Final Pass, distilled model support, and custom resolution.
V2 — Added full built-in Video Editor (RTX Super Resolution + RIFE to 48 FPS), improved LoRA suite, better audio sync, and major stability enhancements.
V1 — The original production-ready workflow. Dual-pass system, strong audio integration, excellent low VRAM optimization (12GB cards), and live previews.
Disclaimer:
This workflow is provided for entertainment, artistic, and creative purposes only. It may not be used for any illegal, harmful, non-consensual, or malicious activities. Please use it responsibly and respect all applicable laws and ethical guidelines.
— MrXin (May 2026) 🔥
Description
Bug fixes:
Changed the Dual Clip Loader device from “CPU” to “default”.
This restores the generation speed from slow back to normal.
FAQ
Comments (153)
Being able to have a section for Audio Input would be fantastic for music videos and the likes. If you do go ahead and do that for us, make sure to add the "audio_start" option so we can choose when in the audio file the generation starts at! :D
Great workflow.
How good is this workflow for anime?
Pritty good if you ask me :p
im using easy install but i cant find the ComfyUI.bat. that is talk about in caption
its in the folder "...\ComfyUI-Easy-Install\Start ComfyUI.bat"
Problems with installing ComfyUI-LTXVideo. Using Stability Matrix I had to downgrade the python package for Korina from 0.8.3 to 0.8.2 (3 dots > Python Packages)
Then disable Sage Attention. WF then worked Using 10Eros fp8
maby a prompt enhancer node for v7?
in my workflow is the node "🎈VRAM-Cleanup" with two fields, "offload_model" & "offload_cache". I don't know why you got something different.
@MrXin Well, the manager installed it. Maybe you have an older version? Can you tell me which one?
thanks. your workflow is the best right now. how to increase quality anyways? iam using 16gb vram. mybe it can increase the quality a little bit.
Resolution is the main thing that affects quality, it even affects audio quality for some reason. always use max resolution if you want quality. Only use lower resolutions for testing prompts.
Bro, you brought LTX closer to the benefits of WAN, while also preserving the benefits of LTX itself! You're great!
t2v version ? :)
is it normal if it takes 40 min?
for 30 sec 1280*736 with 2 extra loras and a long ass prompt
Yes, it's normal for 30 seconds at full resolution with lora.
The workflow works amazingly. However, I am experiencing an issue of a metallic hissing noise throughout the length of the video. I have dialogue and its really good but the metallic noise is constant. Any input would be appreciated
You could lower the audio volume from 0 to -10.
Hey, can you help me with your workflow? For some reason, after around 2 seconds the video starts becoming completely different from the original image. The first 2 seconds look good, then the whole scene changes like the model forgot the original image. I already tried changing the LoRA up/down parameters and different generation settings, but nothing seems to fix it.
I encountered the same problem. I’m not sure if it’s due to the prompt wording. After trying to change the prompt, although the scene changed instantly, the characters’ clothing and general appearance remained the same. It seems I need to keep researching this further.
I just tried it, and indeed it’s due to the prompt wording. The system will only try to maintain consistency between the character and the environment if the prompts are very precise.
V6.1 increases generation time from 300 seconds (V5) to 400 seconds (V6.1) on my 5070ti....
Seems also slower on my 5060Ti with 6.1...
There a 4 extra steps in the first pass, and the nag node also make it slower. You can disable this.
@MrXin Ctrl+B the Negative prompt window, right ?
Have a issue with a costum mode, this one failed to load to load, any one know how to fix it?
ComfyUI-LTXVideo (1)
LTXVSpatioTemporalTiledVAEDecode
I had the same problem today with version 5 of the workflow. I tried reinstalling and performing various troubleshooting steps, but nothing worked, even after searching for the node to see if it was there, but it's completely gone.
i too
On my end, it was an error fetching the Git repository of ComfyUI-LTXVideo.
My workaround was to download the repo as zip on their Github (Direct link: https://github.com/Lightricks/ComfyUI-LTXVideo/archive/refs/heads/master.zip) and extract it in "ComfyUI\custom_nodes\ComfyUI-LTXVideo".
(Edit: I'm using portable ComfyUI, so I also ran a "python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-LTXVideo\requirements.txt" to make sure all dependencies were installed afterward)
looks like the error is to do with ImportError: cannot import name 'pad' from 'kornia.geometry.transform.pyramid'
fixed thanks, Downgrade Kornia
with
\python_embeded\python.exe -m pip uninstall -y kornia
\python_embeded\python.exe -m pip install kornia==0.6.12
@beast18 same error, but
\python_embeded\python.exe -m pip uninstall -y kornia
\python_embeded\python.exe -m pip install kornia==0.6.12
in my case not solvng problem,
can you say what cumfyui manager settings you have, DB, channel, etc?
Personally, I asked Gemini for help. After searching for a while, he told me to go to Custom nodes, ComfyUI-LTXVideo, and find the "pyramid_blending.py" file and delete the pad function on line 6 or 7. So I deleted it after many, MANY attempts. And finally, it worked.
@fatestaynighthyoga84848 thx a lot! Now its working.
@fatestaynighthyoga84848 Thank you. No idea what I am doing for trouble shooting but this solved it after a few frustrating hours of trying to do it myself.
the final out put is more blurry then the first
I love your workflow. It's giving me some of the best quality of all workflows I tried. What would be awesome would be support for a last frame or even LTX director.
I get worse results the same exact loras, prompt, img, etc. with v6.1 compared to v5. The video becomes much more blurry. I've tried adjusting most of the settings but I can't find what's causing it.
I also very rarely get usable results unless I use manual sigmas.
Try v3 eithout final pass it works i think every newer version just gets more broket at least for my rtx5080
Sorry to be so blunt, but I am super sick of workflows that NEVER work. Do you people really test this workflows with ORDINARY normal ComfyUI? I am sick and tired of nodes that REFUSE to work, even fucking installing them manually following every FUCKING instruction on their github account. Jesus! I lost count how many times I install ComfyUI from scratch to run a worklow THAT NEVER WORKS.
If it's that consistent of a problem, it sounds like a you problem.
im shit but im willing to help u i fell yaa this is the first workflow i make work so know the felling
got to agree. I also dont know why it is so hard to create some sort of required tools pack that for sure work, than to have manually find every fng custom node and model which have just a different file name, etc etc . Id rather download 50 GB that I am sure that it works, than having to download all those loose files that you dont know for sure if the files are correct.
I am doing it with ai. I mean instaling things. I ask gemini or grok when I dont konw, or dont like to read to much. copy them eror code, describe the situation, like where is what instaled, that i use stability matrix, what is required and they navigate me or find another way. its not that bad. With this workflow v5. I had one major problem. The ltx audio loader could not see the model file no matter where I put it or renamed. I ended using a copy of the ltx video loader instead, for loading the audio model as well.
@krsny1326 i do that too but the bigest help for me was confyui easy istall it made sageatention work
I used claude to help with the difficult nodes. My problem is I don't even know how to generate anything because the example mp4 files are just to make the video bigger.
@Sasori7 i tougth it woude up scale it and make it cleaner
Your earlier workflows worked without much issues, but this version (with NAG enabled) fails on that specific part :
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
!!! Exception during processing !!! 'function' object has no attribute 'to'
Traceback (most recent call last):
File "D:\tmp\ComfyUI\ComfyUI\execution.py", line 534, 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 "D:\tmp\ComfyUI\ComfyUI\execution.py", line 334, 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 "D:\tmp\ComfyUI\ComfyUI\custom_nodes\comfyui-lora-manager\py\metadata_collector\metadata_hook.py", line 165, in async_map_node_over_list_with_metadata
results = await original_map_node_over_list(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\tmp\ComfyUI\ComfyUI\execution.py", line 308, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "D:\tmp\ComfyUI\ComfyUI\execution.py", line 296, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "D:\tmp\ComfyUI\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\tmp\ComfyUI\ComfyUI\comfy_api\latest\_io.py", line 1826, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\tmp\ComfyUI\ComfyUI\custom_nodes\ComfyUI-KJNodes\nodes\ltxv_nodes.py", line 475, in execute
diffusion_model.caption_projection.to(device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'function' object has no attribute 'to'
my dumb ass keep using Start ComfyUI.bat instad of Start ComfyUI SageAttention so if u read one of the comments i put put in mind that
A support for promt relay would be awesome.
Still, very impressive workflow - thanks for your work!
If I Enable RTX in Image to Video mode it should upscale the video after generation right? Currently it ignores upscaling part completely for me.
Thanks fo a workflow.
you need to upscale it manually after
Why is it a person can't generate an image within the same workflow? what is a good text to image workflow to feed this. it's a real stopping point.
Because often with image generations, you have to try a few to get the image you want. There are often a lot of failures, some requiring inpainting, etc...
Now imagine every single time you try to generate an image, you have to wait several minutes instead of 20 seconds. It makes failures that much more costly, with results you don't like as much. That's why people generate the base images they want first, then go into video generation and use the images they like the most.
BUT, nothing is stopping you from pasting an image workflow into this one, and have that drive the video. All you do is connect the node that goes to "Save Image" at the end of your image generation, into the node that takes the input image in this video generation workflow. Type the image prompt for the image portion, and the video prompt for the video portion. Bingo, you get to waste a lot of time on something that doesn't actually save you anything.
Alternatively, you can just do T2V with Sulphur, and skip the image generation. That would be faster, and you'd have just as little control over the result.
@Jellai the issue is that there is not a set of instructions on how to use it. When I do as you say and download an image/vid with the workflow in it, the workflow is set for just upscaling, not how the image was generated. Everything else is turned off.
@Sasori7 oh, you don't have to download anything then. ComfyUI comes with plenty of text-to-image workflows that don't bother with upscaling. Look in the Templates section.
One of the dumbest things, and I don't know why dingbats do this; it's not just this guy, but WHY DO YOU USE A MODEL THAT YOU DOWNLOADED A 2ND TIME!?
ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3(1).safetensors
So fking annoying.
To be honest the official eros workflow is way way way way better
Agreed. I tried this out with the exact settings and default models and I got junk. The official WFs are an absolute mess, but my first run (after scratching my head over the loading scheme) was what I was hoping to see. Still some troubling artifacts in the upscaler though, better to generate in one pass. Eros beta has been working spectacularly for me, so the bar is pretty high already.
I'm sure I've been guilty of posting workflow versions with bad defaults, maybe that is the case here. But I wasn't really able to solve it without deconstructing it completely.
When you get junk with this workflow, you do something wrong. I get excellent results. The official workflows are for sure nice but with only 12 GB of VRAM they constantly get out-of-memory errors, which is not surprising. This is where this workflow shines.
It's an I2V workflow but it doesn't seem to use loaded image at all. How do you use the image loaded as start image?
Very good workflow that's giving impressive results. Do you plan to release first-last frame version?
SAD, LTX can't maintain the size of penis like wan2.2
no matter how big it is ,it always shrink at 8-10s then it won't turnning back?
is anyone know how to fix it ?
or this 24second video generation ,joint piece to piece?
and low definition works better than high definition....
and low definition less distortion
Maybe someone can help. In this workflow, for some reason, this node (IMPORT FAILED ↗ ComfyUI-LTXVideo) doesn't work. I'll fix it every now and then, but I can't figure out exactly how. It seems to be a problem with different versions of ComfyUI and the node.
When updating ComfyUI, for example, I select the nightly version and click update—it says it's already updated.
If I select the Stable Version and click update, I get the message: To apply the updates, you need to RESTART ComfyUI. Failed to update ComfyUI.
Has anyone encountered something similar and can help?
you need fix import in ComfyUI/custom_nodes/ComfyUI-LTXVideo/pyramid_blending.py
use
from torch.nn.functional import pad
from kornia.geometry.transform.pyramid import ( PyrUp, build_laplacian_pyramid, build_pyramid, find_next_powerof_two, is_powerof_two, )
Really great workflow! Any chance of adding a single pass toggle in a future update? Thanks!!
Has anyone encountered videos that work (somewhat, a little blurry) for a few seconds then go berserk with fog/blurriness like the whole thing is smeared with vaseline? All nodes (except RTX, I'm on blackwell), installed, all models correct.
yeah im on rtx 5080 and still trying to make this work , unfortunately this workflow is done for 3060 - i wish there was a blackwell version
make sure you have downloaded the loras. it will run even if you haven't added them, but if they are not there it goes beserk. it took me ages to work this out but this should fix it
Hey everyone, I have a question. I'm running into a problem using this workflow. When generation is complete in Video Combiner, the quality is good, and I usually save the final result from there. However, the quality in the folder where Comfy saves the final video is absolutely terrible, so I have to stop and save it after each generation. Is there a nuance to the workflow or something? Can anyone help?
this need three dstill lora？？
I am using the Eros model with distil loras with this workflow, 1280 res 12 steps. Im getting really hard swirling noise in the final output. Every moving pixel is heavy grain like film grain, really weird noise. It looks aweful. The example videos here dont have this noise, why is this happening?
me too
make sure you use 384 lora its in the tips int the workflow
@polfa757 it has nothing to do with the distill, i found out by now that its just the Eros model itself which does this. It sjust how it is, its not as clean as other LTX models. Cant do anything about it. You can try to somewhat counter it by cranking up the res but its still very visible even beyond 1920x1080 on fine details like hair and stuff. Im pretty sure the example videos from OP are not generated with Eros but with the normal LTX transformer.
So many problems with v6.1 workflow for me . Ill stick with v4 right now
@Phac123 only saying as it fixed it for me
@polfa757 then you dont know exactly what i mean, because its not fixable.
I'm having an issue with this workflow. For some reason, after the final pass / upscaling, I get noticeable ghosting in the video. It looks like the same faint video is playing in parallel with a delay of about a second. I'm using this workflow in T2V mode after some modification.
Node 'ID #806' has no class_type. The workflow may be corrupted or a custom node is missing.
Node 'ID #806' has no class_type. The workflow may be corrupted or a custom node is missing.
我的工作流不知为何10s-comfy-nodes节点明明已经安装了，但依然显示缺失节点包
Any way to turn off NAG (Negative prompt)?
For people who use blackwell/cant get it to work - its a tough workflow as its no optimised for your hardware but after trial and error this is what works for me :
blackwell master porteable comfy UI
remove VAE dual loader - i used individual loaders for audio and video VAE like in V3
chunk setting 4 and 4096
node 149 settings:
spatial_tiles = 2
spatial_overlap = 8
temporal_tile_length = 64
temporal_overlap = 4
last_frame_fix = True
working_dtype = bf16
sage attention disable in the workflow but added flag to my bat to enable it so launches with it anyway.
for final pass make sure you use 384 lora as per tip in the workflow
for RTX upscaler watch youtube video from prompt mastery type RTX and its there.
If your upscaler giving you hard time dont disable it as per tip in the workflow you will break it . open your bat flag and remove --no cache flag. rest of the settings you can leave as is but image is super inportant- hope this helps someone i had to work it all out myself..
MrXin EROS V6.1 FLF I'm very bad at building my own workflows. Is it possible to add First Frame and Last Frame here and how to do it without breaking anything?
It work perfectly for me, and the results are very good, thanks!
though, I have some issue with some kind of CUDA errors earlier but got it fix with this tag to fix it: --disable-smart-memory --disable-pinned-memory --cache-none
This might be a stupid question.. How is it possible to run this workflow on a 12GB VRAM card when the models are massive? 20GB+.
models are cached in your system RAM, VRAM is used to actually run the model so it doesnt have to hold the entire size. I dont know if thats technically exactly how that works but that is why you can do that. That said you do need like a shitload of ram lmao, I have 96gb.
take care to have 32gb DDR or more. Also set up a swap file on your fast nvme hard drive.
It runs well even on rtx 3060ti 8gb. I have 32gb ram and 96gb swap file.
@wemite Can you tell me how long your generations take?
@dreadfulpirate 4-5 mins for 15 sec 720x960. ofc with sage attention. also I tried it on linux. it faster for like 15-20%
Very nice WF, I use it with Dasiwa v3 LTX model it's even better than eros imho.
I'm using GPU 1070 so I cannot use SageAttention. How can I disable it? Thank you!!!
Quiet Audio Issues? Anyone else getting super low audio outputs? I changed it to 10 and it's still very quite. Never had this problem before. Do I just crank the audio volume up to 20?
It would be great to integrate LTX Director node :)
I am having problem with LTXVSpatioTemporalTiledVAEDecode not being installed.. I tried so many times to install missing nodes.. but never work
had the same issue - try this.
install ComfyUI-LTXVideo
open custom_nodes\ComfyUI-LTXVideo\pyramid_blending.py
remove line 13 where it says "pad,"
add a new line 14 (after the ")") "from torch.nn.functional import pad"
restart comfyui
I keep getting an error saying that the comfyui_nvidia_rtx_nodes are missing. However they are installed using the manager. any advice? everything else seems to be working it is just this one node.
It doesn't run on ARM64 if that is what you are using.
Hello, amazing workflow, thank you.
My only issue is on I2V there's a lot of identity drifting. Any idea on how to fix it?
Really great work. Would you consider adding prompt relay and video extension capabilities to give more control over timing and the ability to make videos longer than 30 seconds?
RuneXX on huggingface has some good workflows with video extension: https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main/Video-2-Video/Extend-Any-Video
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 256.00 MiB. GPU 0 has a total capacity of 15.91 GiB of which 189.75 MiB is free. Of the allocated memory 15.09 GiB is allocated by PyTorch, and 291.11 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) This error means you ran out of memory on your GPU. TIPS: If the workflow worked before you might have accidentally set the batch_size to a large number.
Need some help with a missing node pack for RES4LYF. Even though it is installed in my cusotm_nodes folder, I get an error saying it isn't installed yet. I've removed it, reinstalled it, and even downloaded a copy from Github that gave me the same error. I'm not sure what is going on, but seeing as how it's my last remaining problem to fix, I sure would appreciate the help.
I have the same problem .. Not sure how to get that to work.
My custom node I haven't installed is called "LTXVSpatioTemporalTiledVAEDecode" so far i have had luck by using the terminal to git clone from git hub (I used to have 7 missing nodes). Then moving the files and folder from one custom_nodes folder to another... for some reason I have two. The file structure near the models that comfy installs seems to be the correct one. the file structure the console points to not so much. I can only assume I am using the wrong version of ComfyUI-LTXVideo. I dont know if I am helpfull but at least we are both suffering. oh I worked it out. It only works with nvidia. I have it installed but it wont run.
Guys, can somebody help me?
I'm trying to combine IC-Control Lora with MrXin workflow, and when I add the Video IC- control Guide node between LTXVConditioning Video and CFGGuider First Pass, the video doubles in size and the rest of the video is filled with noise.
and it doesn't matter how many seconds the control video I attach is - 5 seconds - it already turns into 10 on the first pass. 20 seconds - on the first pass, it turns into 40 - 20 of the character moving correctly + 20 of just noise.
does anyone know how to add LTX audio to a video?
I gotta be doing something wrong. No matter what my prompt is, the audio is terrible. So much static.
Bro, this workflow is just amazing, its easy to use. With my 8Gb VRAM and 32RAM its take 560s for 10s video 1080x600. Who needs a WAN now XD? Thank you for the work done.
Has anyone experienced this problem with ComfyUI_NVIDIA_RTX_Nodes ?
Traceback (most recent call last):
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\nodes.py", line 2212, in load_custom_node
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 999, in exec_module
File "<frozen importlib._bootstrap>", line 488, in callwith_frames_removed
File "D:\ComfyUI-Easy-Install\ComfyUI-Easy-Install\ComfyUI\custom_nodes\comfyui_nvidia_rtx_nodes\__init__.py", line 2, in <module>
import nvvfx
ModuleNotFoundError: No module named 'nvvfx'
Same issue for me but it's only used in the video editor
Manually install the RTX node for your setup. You'll see in manager import failed.
This is the kind of problem that an LLM will fix for you in two shakes. It says 'no module', the module is missing. Any free LLM can give you the commands for manual installation with troublesome packs.
For what it's worth, I had this weird issue where sometimes the output was a single frame, from what I remember it had something to do with the dimensions not conforming to something or other. Tensor flops multiplexorizer whatever, didn't throw errors though. That was with 4090. When I use it on the other machine with a 5090 it runs with no problems. It was never a huge deal for me, as there are alternatives, so I didn't have to hunt down exact causes, but I can at least confirm odd behavior from the node, even without errors.
Unfortunetly anything that requires the ComfyUI-LTXVideo is a no go in ComfyUi for me. Never installs. Just errors. Us AMD users are DOA on lot of these nodes.
Any more updates ? LTX 2.3 10Eros v1.2 is out. I tried with your 6.1 workflow but it doesnt work well most of the time...
Is it ok to have this error:
VRAM Cleanup "Required input slots have no connection feeding them."
wf runs fine but maybe i need to connect it to something?
I never connect my nodes together, in any of my workflows. It's blazing fast. 0 VRAM usage, it's friggin incredible. You're probably missing the impact pack, as the only thing that feeds the purges is a dummy. Some comfy version/OS/dependency combos are crazy leaky and hang to stuff that should not persist. Pinned memory, smart memory, floppy flags and whatnot. Just keep an eye on your resource meters. RAM and VRAM should drop after a run, (VRAM depending on unload settings). If you don't have problems you can ignore it. But it's good practice. You can also use the 'filenames' output of the VHS combine node to trigger the purge instead of that dummy node. I like to do a RAM cleanup triggered by the incoming image and a VRAM purge after the combine.
How well would a 5060 ti 16GB do?
if u set up too high resolution it will crash
Nice setup, I've always found the beta version superior but this does give decent results for v1. I would point out one thing- you've got an error in your frame count logic. #42 of the subgraph uses a*24+1, (where a is duration) which means that the duration will only be valid when 24fps is used in the promoted widget. The expression should be (a * b + 1), where b is the fps float. I add a bit more myself, to round to the nearest compatible count. But that's the base expression, in case anyone needs to fix it.
Oh also the fast bypasser that switches between full and distilled models does not disable the distillation LoRAs. They should be in the same group as the loader for the full model, or they will still load when using a distilled model. Double distillation.
how to disable sample (file first.mp4) and audio i dont need it (first_audio.mp4, final_audio mp4)
Settings > VHS > Two top options for keep and save stuff, tick them off.
no matter what ltx workflow i try , all give me " RuntimeError: The size of tensor a (10736) must match the size of tensor b (1438464) at non-singleton dimension 2" error granted its not always the same numbers not without a doubt its the ' RuntimeError ' error
You are using Models that are not compatible with your Text encoder and or your VAE...Check carefully what works as designed. Best to ask ChatGPT with your exact Setup...
@Ai_Unleashed oh right okie, thank you, ill have another look at the workflow and its requirements
@djacid Did you ever solve it?
I finally found it on my side, I needed to uninstall smznodes, without further changing the workflow, is now working.
@soiasul i didnt but ill give that ago, im waiting on a new GPU atm should be here in the next hour or so , once it arrives ill have another look, thank you for in info
I come from using Wan 2.2 and I just switched to using LTX 2.3 with this workflow. I have a question: I see there are two models a 10 eros model and a distilled model. Are they used together at the same time, or is it either the 10eros model or the distilled one?
I input a 1280*720 image, keeping all other parameters unchanged, but the generated video quality is very poor. What settings are incorrect?
Make sure the aspect ratio in the resize image parameters matches the input image resolution exactly. I had the same problem.
I put in an image, but no matter what I try the video loses the face of the input. Is there a fix for this?
have a look at Best FaceID v1.0 lora, works absolute magic for me
@devilmastah128 can you point us to this lora? Search seems to be broken right now...
@meropewantsstls There were 2 wf's included with the lora. The first one's results were junk. The second one (biggger one with two passes) made the face have a slight resemblance to the character, but came nowhere close to holding the characters identity.
Is there a way to control the denoise strength with this? For some images in the I2V, the initial frame changes quite a lot from the original image
Is there a way to increase the video size to more than 30 secs?
Yes it is. Right click on the "Video Length (seconds)" Node, then Properties and change the value "max" to any number u like
Missing the node: LTXVSpatioTemporalTiledVAEDecode
Despite having ComfyUI-LTXVideo node installed via the manager. Any ideas?
is the node pack actually importing? if not you might want to look at why it's not; it should give you errors in the logs - like missing a dependancy
Downgrade Koira to version 8.2. There is something that is borked with v8.3 and up.
i would love to see this updated
Man, I really need some help here because I'm completely at a loss. I've tried getting into LTX three different times since launch, and every single time, it completely freezes my system. I just gave it another shot, and it locked up right on the first pass. I even tried running it with these arguments: --use-sage-attention --lowvram --reserve-vram 6 --preview-method none --disable-xformers --disable-smart-memory This time the freeze wasn't permanent, but it just threw an OOM error instead. I honestly don't even know where to ask for help anymore. My Specs: OS/Env: WSL GPU: RTX 5090 (32GB VRAM) RAM: 64GB System RAM (58GB allocated to WSL) I'm honestly ready to just give up on LTX at this point. Any ideas?
Have you tried a complete reinstall of comfyUI? Every couple of months I need to do a complete fresh reinstall or things like these start happening. It's a pain, but that's just comfy.
This is Kind a Good , But in Video Editor I am getting OOM error even though I have 16GB VRAM and 32GB RAM in my system.
Could you tell the settings that I can change to solve this OOM.
I am using v6.1 and the videos become fuzzy and audio becomes staticy/garbled. Does anyone have and recommendations or advice?
I will greatly appreciate it.
Open the console and look for [Power Lora Loader] Lora "..." not found, skipping lines. This is the big one — the workflow depends on two distilled LoRAs (ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe first pass, ltx-2.3-22b-distilled-lora-384-1.1 final pass), and rgthree silently skips them if missing — ComfyUI's missing-model checker doesn't scan LoRA loaders, so you get no error. Running the 8-step manual-sigma schedule without the distilled LoRA = soft/fuzzy output every time. The condsafe one is on TenStrip's LTX2.3_Distilled_Lora_1.1_Experiments HF repo, the 384-1.1 is in the official Lightricks/LTX-2.3 repo.
Swap the spatial upscaler to v1.1. The workflow references ltx-2.3-spatial-upscaler-x2-1.0, which has known artifact problems (worst near the end of clips). x2-1.1 is in the same Lightricks repo — just re-point the Latent Upscale Model node.
For the garbled audio: make sure the audio VAE is the right file (LTX23_audio_vae_bf16 from Kijai/LTX2.3_comfy) and try running it at fp32 instead of bf16 in the VAE loader node — audio decode is much more sensitive to precision than video, and static/crackle is the classic symptom.
Also make sure you're pairing the v1.1 distilled LoRA with the model — v1.0 LoRA on the current base causes stiff motion and color drift.
@Z1LoX - I appreciate your help. I was going to post that I found the problem. It was the aspect ratio did not match the input image. I felt like an idiot once I realized that. After that, it works like a charm.