Update 2026-05-13: LTX 2.3 All-in-One v3.0 workflow published
VideoFlow LTX 2.3 distilled 1.1 All-in-One v3.0
New Features:
Text-to-Video Support with a pre-configured set of LoRAs for creating photorealistic videos.
Image-to-Video Support for both first and last frame.
Optional Audio Integration (audio-to-video) with the ability to extract voice from recordings (file or recorded clip) to remove background noise.
Consistent Character Voice through voice cloning with just a 5-second reference audio (file or recorded clip).
Video Filters for adjusting brightness, contrast, saturation, sharpness, blur, and enhancing edges and details.
Film grain for a cinematic or analog effect.
50fps Support via frame interpolation.
Improvements:
Now using LTX 2.3 distilled 1.1 resulting in better emotions, movements and audio.
Faster, less memory-intensive color correction.
More explanations and guidance integrated.
Fixes:
Audio and video are now always perfectly in sync.
Resolution input (video dimensions) for image-to-video generation now works properly.
Update 2026-04-14: LTX 2.3 I2V workflow updated
VideoFlow LTX 2.3 distilled I2V v2.0
VideoFlow 2.0 is here, bringing major performance upgrades, better quality, and more flexibility to your workflow.
Key Improvements:
Much Faster Generation: Thanks to improved samplers and schedulers, videos generate approx. twice as fast compared to Version 1.0.
Higher Quality Output: Despite the speed boost, image quality, audio quality, and prompt adherence have all been significantly improved.
Flexible Model Support: You can now freely choose between multiple model types:
Checkpoint
GGUF UNet
Diffusion model
Optimized for Low VRAM Systems: With GGUF support, VideoFlow now runs much more efficiently on systems with limited GPU memory.
Optional Sampler Preview: Disable the sampler preview to further reduce generation time.
Improved Usability: Additional guidance and hint texts help you get the most out of the workflow.
Update 2026-03-15: LTX 2.3 I2V workflow added
VideoFlow LTX 2.3 distilled I2V v1.0
This workflow provides an easy-to-use image-to-video solution for LTX 2.3, designed to work seamlessly with the distilled LoRA model. It focuses on high-quality, realistic output, with the first-stage scheduler's sigma values finely optimized for best performance.
Subgraphs are used to keep the main workflow streamlined and easy to navigate. A live preview is displayed during generation, allowing you to monitor progress and stop the process early if desired. Additionally, the first-stage video can be decoded for quick previewing. This feature lets you watch a lower-resolution version of the final video and cancel immediately if the result doesn’t meet expectations.
As the distilled LoRA already delivers impressive quality in the first stage, you can skip the second stage entirely if your hardware has limited performance. An optional color-correction node is included to compensate for LTX’s tendency to introduce subtle color and lighting shifts, ensuring consistent visual quality.
Update 2025-08-24: Wan 2.2 I2V workflow added
VideoFlow Wan 2.2 I2V v1.0
VideoFlow is now fully optimized for Wan 2.2. It supports resolutions from 480p up to 720p, with the option to upscale smoothly to 1440p at 32fps. The process is accelerated by integrating Lightning LoRA during the final two-thirds of the generation steps, ensuring faster results without compromising quality. Importantly, Lightning LoRA does not influence the initial generation steps, preserving natural and fluid movements throughout the video. SageAttention with Triton is supported but not required. Instructions on how to set up and use the workflow are included within the workflow itself.
VideoFlow Wan 2.1 I2V v1.0
This image-to-video workflow is designed to generate smooth, realistic videos at 32 fps with a strong emphasis on fast, high-resolution output. At least 16 GB of VRAM is recommended for optimal performance. For additional speed improvements, you may also install SageAttention and Triton, though these are optional.
It's fast 🚀!
Sample videos were rendered at 768 × 1152 resolution and 16 fps, consisting of 81 frames, each video taking about 6 minutes to generate. The upscaling and frame interpolation to 1536 × 2304 resolution and 32 fps took approximately another 6 minutes on an RTX 4080 with 16 GB VRAM. Lower resolutions render even faster.
Key configuration for the sample videos:
Video model: Wan2.1 SkyReels V2 I2V 14B 720P
LoRA: Lightx2v
Steps: 4
Sampler: dpmpp_sde_gpu
Scheduler: beta
💡Comprehensive usage details and instructions are provided within the workflow itself.
Sample images for input were created with my PhotoFlow workflow.
The download of the workflow contains all sample videos, including the input image with its own workflow, the initial generated video and its upscaled counterpart, allowing for convenient side-by-side comparison.
Leave a 👍 if you like the workflow 🙂.
Description
VideoFlow 2.0 is here, bringing major performance upgrades, better quality, and more flexibility to your workflow.
Key Improvements:
Much Faster Generation: Thanks to improved samplers and schedulers, videos generate approx. twice as fast compared to Version 1.0.
Higher Quality Output: Despite the speed boost, image quality, audio quality, and prompt adherence have all been significantly improved.
Flexible Model Support: You can now freely choose between multiple model types:
Checkpoint
GGUF UNet
Diffusion model
Optimized for Low VRAM Systems: With GGUF support, VideoFlow now runs much more efficiently on systems with limited GPU memory.
Optional Sampler Preview: Disable the sampler preview to further reduce generation time.
Improved Usability: Additional guidance and hint texts help you get the most out of the workflow.
FAQ
Comments (46)
we want videoflow svi v2 pro workflow for wan based on ur three sampler method ,its good and best method for wan . and ofc ur 3 sampler settings r best .
does this workflow support audio generation ? because claude said this cannot generate audio.
LTX 2.3 generates videos with audio so it is supported in my LTX 2.3 workflow. You can see and hear this for yourself in the examples here. All you need to do is turn on the sound for the videos, as they are muted by default here on Civitai. The Wan workflows are without sound.
RTX 4080 16gb VRAM & 64gb RAM, 22b Dev UD Q4KS GGUF, Gemma fp4 mixed with text projection bf16, regular undistilled voice/video VAEs, x2-1.1 upscaler, 8 steps @1 mega pixel skipping first stage:
3 second test video with distilled lora took 187 seconds with reasonableish quality.
Second run same settings but bypassing Lora took 105s basically didn't work - no movement.
Third run same settings lora at 30%: 161s didn't move much but did work.
5 second video back to 60% lora: 181s reasonable movement. (I noticed VRAM was at about 84% for much of it, but I do have a vram setting in the launcher extraargs.)
10 second video: 242s.
Would any of the RTX/DynamicVRAM nodes help make this faster or better?
Actually I forgot to add at least one node. My fault. I will give an update on this issue today.
@Phraxas I am still figuring out the best settings and node configurations. Is there a special reason you chose the dev instead of the distilled GGUF file? The distilled one has a very similar output, at least in my tests for image-to-video. This would safe some RAM/VRAM because no distilled LoRA is needed anymore.
@ai839 I was using other workflows that were throwing CUDA errors at me when I used distilled and I have absolutely no idea why, so I sort of defaulted to using that model which ended up working fine on yours. I can give a few other models a try. I'm also going to try pumping it to the normal 2.1 megapixels for 1080p.
@Phraxas Another guy just tested successfully 2.1 megapixels. 10 seconds. 2x latent upscaler. 12 GB VRAM (RTX 4070), 32 GB RAM, no VAE Decode Tiling. Model: ltx-2.3-22b-distilled-Q4_K_M.gguf
@ai839 Yeah, I ran a few times with the distilled and it's definitely faster and works just fine. I'm at about 200s for 10s video at 1mp with 1 lora active (not the distilled one; just a regular lora). Ran it again with 2.1mp and vae decode tiling off and it completed in 440s.
Keep in mind my launcher vars are "--disable-cuda-malloc --reserve-vram 4" I'm not sure if that's a problem or helping.
This was just a quick test - I still had a few things open for a project I didn't want to close out, but it wasn't much RAM being taken by it and no VRAM.
None of this is limit testing. If you have anything you want me to try to push it, let me know. I might give tile size a bump.
Low-VRAM: HELP NEEDED🙂
Anyone with a system that has limited VRAM (16 GB or 12 GB) and is willing to try a few tweaks in the LTX 2.3 workflow to find the best settings can contact me via Civitai Chat. I appreciate any help.
Other workflows show the preview.. but not this one.. any setting I might be missing?
The LTX 2.3 v2.0 workflow has an extra option for it in the Stage 1 Sampler and Stage 2 Sampler node. The second last option. Check if "Live preview method" in the ComfyUI settings is set to "auto". Not sure if this is needed here but I have it set to auto.
I liked it very much, thank you for that. Are you planning to make a text2video version?
Not sure yet, maybe.
Update: I'll be releasing a text-to-video workflow for LTX 2.3 soon. There will also be an update to the image-to-video workflow. I haven't decided yet whether to offer two separate workflows or a single one that supports both.
The output quality is already very promising, and it looks like I've fixed the issue with audio going out of sync in longer videos. I'm not sure exactly when everything will be ready for release. I expect it will take a few days, not weeks.
Hi Dev, great workflow!
I'm experimenting with it a bit, and while I understand that it's designed for distilled, when I do use the undistilled fp8 model the results are just blurry fuzz with distorted audio—there's no instructions for changing settings for base / fp8 models so I'm wondering if you might comment if there's something I'm overlooking? I was thinking maybe CFG, but I'm not finding it easily within the subgraphs or settings. Hoping you might enlighten me or have a clue as to what is happening? To be clear all distilled models are working great.
Yes, I enabled the distilled lora at 0.6 for the fp8 model and read your entire detailed instructions, so I'm hoping that you can help for when I want a bit more emotional / performative output with fp8 non-distilled checkpoint.
If you have any questions for me I'm happy to answer, if you think it will help. Any insight is appreciated.
Great work!
TIA!
What you describe happens if the distilled LoRA is not applied. Double check that the toggle is not switched to off and the LoRA is really selected (maybe there is only the LoRA name from the default and your file is named differently). Select the LoRA file again in the LoRA loader just to be sure. non-distilled models work perfectly with the LoRA. The CFG needs to be at 1.0 all the time. The sheduler and sampler in this worklow do not fit for non distilled generations at higher CFGs.
Actually my outputs are more emotional / performative with the distilled models. A lack of emotions is often caused by bad prompting. ChatGPT is quite good to assist with that!
@ai839 Thanks for your response! Yes, I forgot to load the LoRA manually—as it's stored locally in a subfolder rather than the LoRA root folder. Confirmed working!
Hello Dev,
Another question. If I want to use resolution (smaller side, i.e. 720) rather than megapixels (2.1 or 1.1), how do I force that? I see some boolean nodes that are minimized in the subgraph, but I'm not seeing an on/off switch that is apparent.
I was hoping that it was driven by typing the res number in, however if I zero out the megapixels (value=0) I get an error stating my resolution value must be set to '0', and the video won't render.
And If I leave the megapixels number (say 1.1 or 2.1) AND type the smaller size resolution value (720), it seems to just ignore the number and go with the megapixels value (I check resolution and it's not 720, so I assume it's following the megapixel value).
Any insight is appreciated, thanks again!
The resolution option will be used as soon it is not 0. The megapixels don't have to be set to 0 in this case. Exactly 720 will never be applied, because LTX needs all sides to be divisible by 32. If you use the 1.5x upscaler with 720 resolution the smaller side will actually be 672 pixels (because this and 672/1.5 is the closest lower number that is divisible by 32). If you use the 2x upscaler 704 will be the shorter side (704 and 704/2 divisble by 32).
@ai839 Thanks for the response! What about 768? (32x24) And do I need to add a +1 like a lot of other workflows require (divisible by 32 +1, i.e. 769)—if I want to get my numbers right?
@1BadAshB 768 works fine. Wrong numbers will be corrected automatically. You cannot do something wrong. Adding +1 is not needed. The +1 is for the frames, but this is done automatically in this workflow.
Thanks, very interesting workflow! It seems to work fine. The audio is however out of sync by a large margin. Not sure what to change to get it in sync. Using the distilled checkpoint and your latest workflow for LTX
I have the same problem with all models and all other workflows. Seems to be a LTX issue?! Above 10 seconds it starts to get out of sync. Sometimes its better, sometimes worse. If somebody has a solution for this, please let us know!
It helps if the video's length matches the prompt's requirements and if it is completely clear, both visually and audibly, what is happening at any given moment.
@ai839 Ok maybe a common problem? Oddly, when I used the smaller FP8 version, the one around 28 gb or so, I had perfect audio sync. (But that was another WF)
@snobbias124 The question is whether the audio is always in sync with the other workflow. Sometimes it's also in sync with my workflow. There is some "seed luck". Could you share that workflow with us? I'd like to try it out myself. I also tried fp8 and all the other models (fp16, GGUF, distilled and non-distilled models with the lora).
@ai839 I did some attempts using FP8 in your workflow, and generally speaking, is seems to behave more predictably unless I'm imagining things. I did not experience audio oos up to 14 sec, didn't go any further though. My other wf was one of the "official" ones, from the templates (the one with a switch to select between i2v or t2v). It defaulted to FP8 though, and no matter how I tweaked, I could never get the full distilled to work properly (bad video quality). I can't share it since I have messed it up too much during my troubleshooting, and the ones present now from the templates seem to have been updated.
I'm just a beginner, but in my experience, the full distilled LTX 2.3 model requires extreme attention to all settings while the FP8 seems to be more forgiving.
Speaking of tweaking, I can't get triton/sageattention 2.2.0 to install properly using Cuda 13.0 (ubuntu) but I suppose that is to be expected. And the old sageattention 1.0.6 is not recognized by the wf. I'm way over my head here, so many questions :)
@snobbias124 I also tested the official workflow templates and had sync problems, too.
Regarding SageAttention 2 I would ask an AI for assistance. I am using Windows and haven't installed it in an Ubuntu environment yet.
Update: I'll be releasing a text-to-video workflow for LTX 2.3 soon. There will also be an update to the image-to-video workflow. I haven't decided yet whether to offer two separate workflows or a single one that supports both.
The output quality is already very promising, and it looks like I've fixed the issue with audio going out of sync in longer videos. I'm not sure exactly when everything will be ready for release. I expect it will take a few days, not weeks.
Testing v2 vs v1... Not seeing the 2x speed improvements mentioned, even using the new distilled v3 model with the lora disabled as designed. Additionally, the completed video no longer "plays" when mousing over it in the stage 2 video node at the end, like it does in the V1 workflow. I just get a stationary image. Not sure what causes this, but I was able to fix this by clipspace copy the stage 2 video node from your first workflow (v1) into the second one (v2) and reconnecting the noodles.
RTX 5090, 64GB DDR5/6000
Interesting. We have a similar setup. I have an RTX 5090 and 96 GB of RAM. I don't remember the exact length of the video, maybe 10 or 13 seconds in my test, but I have some notes here:
Version 1.0:
Stage-1 sampler: 3:19
Stage-2 sampler: 2:57
Both samplers: 6:16
Version 2.0:
Stage-1 sampler: 1:07
Stage-2 sampler: 1:47
Both samplers: 2:56
Did you change the sampler in the subgraph for Version 1.0? The speed boost in Version 2.0 is due to the fact that *_pp_cfg samplers are no longer used. If you have already made this change, there will be no further 2x speed increase. In this case, sampling will be 20% faster in Stage 1 and 33% slower in Stage 2. The step sizes have been changed from 10 to 8 and from 3 to 4.
I am using the Nvidia Studio Driver 595.79 from March and the latest ComfyUI standalone version (Python 3.13.11, PyTorch 2.10.0, CUDA 13.0).
If you right-click on the video preview in the node, you can click "Pause preview" or "Resume preview". Toggling this option should resolve the issue.
@ai839 The samplers are definitely different in the second workflow, I did take notice of that, just not seeing the speed improvement. I may need to reinstall CUDA, I know it's out of date, but I'm dreading it as it usually breaks something.
edit* : Updated CUDA to latest version. Using your basic prompt used in your sample image for the model, I generated a 10 second video in 192.97 seconds (3:12). Thanks for the reply, I think this seems to have fixed it.
@Minase460 thx for the info. Do you know how long it took before the update? Which CUDA version was installed before?
@ai839 I'm honestly not sure, I believe it was a 12.variant. Gen times before the update were closer to v1, (in the 5-600sec range). I also recently updated to the 5090 from a 4090 so that update was probably long overdue. Having played around with the v2 workflow a bit more, I'm still having some quality issues. Although it's definitely faster, it seems to produce lower quality results, mostly only with the use of LORAs. I know LTX is very touchy with sampler settings and other things when it comes to LORAs. It works acceptable enough given the speed improvements, but I definitely notice quite a bit more artifacting on teeth (common) and other things, and lack of detail pulling the data out of the LORA. It's a struggle always with these models, but this is probably still the best workflow I've used for it. I appreciate the update.
@Minase460 I didn't have time to test LoRAs but these are different ideas I have to solve the problem:
- Switch from euler_ancestral to euler in Stage 1
- Increase steps first in Stage 2 (something between 5 and 8) and if this is not enough also in Stage 1 (10 or more)
- Increase the megapixels (higher resolution can solve teeth issues)
- Switch to euler_ancestral_pp_cfg or euler_pp_cfg (most other workflows use them)
Maybe something of this or a combination increases the quality for your videos.
I guess the biggest problems of most LoRAs is the poor quality of the training data. This is something that cannot be fixed by settings in the workflow.
You need to update your comfyui and also update all the nodes in the workflow, you might get some api deprecation warning, you can ignore them. Also for some weird reason fp16 accumulation does not work well for me, so I disabled it.
You're just amazing! Love your work. The detail!!? Wow.
And dude, your sense of humor. Next level1
You're a fking rock star.
VideoFlow LTX 2.3 distilled I2V v2.0 - DOA, doesn't render on H200 even with all dependencies met. Don't waste your time on this one.
Hi, I don't see the settings nodes for length and resolution in the LTX2.3 i2v workflow. Did I miss something?
They are in the first blue "Stage 1 Sampler" node.
super workflow!
Can you also release a T2V workflow for this as well? This is an amazing workflow btw.
Didn't have time for it until now. Maybe someday...
Update: I'll be releasing a text-to-video workflow for LTX 2.3 soon. There will also be an update to the image-to-video workflow. I haven't decided yet whether to offer two separate workflows or a single one that supports both.
The output quality is already very promising, and it looks like I've fixed the issue with audio going out of sync in longer videos. I'm not sure exactly when everything will be ready for release. I expect it will take a few days, not weeks.
this gives incredible quality on a very reasonable time, however it looks like distilled v3 version gives some very noticeable slowdown in my case compared to v2, and overall i'm getting some slow down with this, very wan-like, anyone else?
I dont know why, but I get good results in this workflow only using GGUF. Other two options are just blurry and audio mess. Dont you know why guys ? What am I missing.