Even though it's called a "Starter Suite", it's not just for learning the basics—you can use it in many different ways depending on you. The reason this workflow package is as large as 500MB is because it includes six sample videos in a fully reproducible form. For installation and usage instructions, everything is explained in detail in the YouTube walkthrough below, so please check that first.
Due to compatibility issues between Node 2.0 and the rgthree nodes, I've removed rgthree from my workflow. It was only being used for seed generation, so it doesn't affect anything, but the workflow embedded in the sample videos still includes rgthree. If you want to try the samples, I recommend using the workflow.png file instead of dragging and dropping the video to open the workflow.
Here, I'll add some notes on things you should adjust depending on what you want to generate, points that were not covered in the video.
[WanVideoSampler: steps]
Adjust the values based on the complexity of the motion. Setting a high number of steps for simple motions with very little movement is just a waste of time.
6: This is the baseline value. Try using it first.
8: Select this when the character's movement is extremely fast, involves rotation, or moves significantly forward, backward, or sideways.
https://www.youtube.com/shorts/V3b8lLm9wF0 Example of extremely fast movement.
https://www.youtube.com/shorts/J8gyHmYRHlE Example of a character spinning repeatedly.
10: Select this for extremely difficult generation tasks, such as drawing the character's back view while keeping the outfit consistent even though it is not shown in the input image. Even then, the results will still depend on luck...
https://www.youtube.com/shorts/l2vxSoJSV3o The part where the back view continues for four seconds is extremely difficult, almost like a nightmare.
[WanVideo Context Options: context_overlap]
Increasing this value makes the motion flow more smoothly and reduces background warping, but it also increases the generation time.
[WanVideo Decode]
In my workflow, videos at a resolution of 720×1280 are decoded by splitting them into 8 tiles. Using more tiles reduces the load, but it also increases decoding time, so ideally you should adjust the settings to match the resolution you plan to generate. If you want a rough suggestion for values that fit your environment, you can ask ChatGPT to calculate them. If you're not sure what to use, the default node settings (tilex: 272, tiley: 272, stridex: 144, stridey: 128) should work fine.
[Lora: FusionX]
As a side effect of its very strong stabilization, it tends to generate similar-looking faces. This becomes especially noticeable in close-ups or smiling shots, so lowering the value or not using it at all can be an option. There are many types of speed LoRAs, so if you find good options or combinations, please let me know.
Thanks to all the model, node, and LoRA creators, and to the community and followers who give me motivation and inspiration.
Description
Initial release. Includes the generation workflow, the pose estimation workflow, and six sample videos.
FAQ
Comments (65)
Unfortunately, that "video_config_animate" file is not included. But I see a big improvement in the person's face. Congratulations.
Oh no, I forgot to include that. I’ll reupload the archive. Thanks for letting me know.
@UncleHooru No problem. I really like the quality of the videos. Did you use an upscale?
No upscaling, no interpolation either. I’ve included the missing file now.
@UncleHooru Indeed, "Phantom_Wan_14B_FusionX_LoRA" brings some changes in terms of color towards the end (making them more pronounced). I noticed that you are not using "relight lora." Is there a specific reason for this, or did you simply forget to add it?
Thank you so good workflow.
Thank you for your support. I really appreciate it 😁
The workflow is capable of easily generating high-quality videos. The only mention is regarding "Phantom_Wan_14B_FusionX_LoRA," which makes changes in terms of color quality. It makes them too pronounced. I tried without "Phantom_Wan_14B_FusionX_LoRA" and noticed slight losses in video stability. The details in the video lost some of their consistency. It is possible to work with this lora, but at a much lower value.
There are also inconsistencies regarding background elements. Some change, others disappear and then reappear.
Anyway, congratulations on the workflow. I am happy when such high-quality workflows appear.
Hello,
I am currently renting cloud GPU instances for AI video generation. My current task involves generating a 20-second video at a resolution of 864×1536. However, the process is taking approximately 3,500 to 4,000 seconds (about 1 hour) to complete, which is quite inefficient.
Could you please advise on any methods or strategies to accelerate this workflow? Your insights would be greatly appreciated.
@huoshuimoji716 There is no strategy, at least for this workflow. Unfortunately, this workflow had a lot of potential, but the fluctuations in the background and outfit are exaggerated. Indeed, the generation time is far too long. I wouldn't mind the long generation time if it weren't for these fluctuations. But unfortunately, I don't think the author of this workflow will improve it. I think he has abandoned it. All we can hope for is that the full model for SCAIL will appear.
@drak0n Haha, can't believe Wan 2.6 Animate actually went closed source. Guess this workflow is gonna stagnate for a while.
@huoshuimoji716 If it takes about an hour to generate a 900p, 20-second video, I can’t help but be satisfied with this model/WF. Unlike image generation or I2V, where you repeatedly generate results and pick the best one, this model—depending on your skill level—can often produce a finished result in a single run, so it can be considered efficient. If you have enough resources(VRAM) to spare, reducing the blockswap value might speed things up?
@drak0n Distortions in the background and clothing can be significantly reduced by choosing Euler as the sampler. The situations where Euler/Beta works better are quite limited. I didn’t fully realize this at the time I released the workflow. When it comes to SCAIL, I believe SCAIL-POSE is the core of it. I hope other dance models will support that format as well. While there are motions that only SCAIL-POSE can analyze, there are also situations where ViTPose or SDPose perform better. Ideally, it should be possible to choose between them depending on the case.
@huoshuimoji716 I haven’t made any changes to the workflow itself, but I feel that the results are steadily getting better (I haven’t posted them due to an issue on Civitai’s side 😅). This is mostly due to improvements in ComfyUI’s memory management and updates to VideoWrapper now allow heavier workloads and higher-resolution generations to go through. Another factor is that I’ve personally figured out better ways to generate base images that look less unnatural once animated. In short, I’ve realized that the workflow itself isn’t all that important having a deep understanding of the model is what really matters. I imagine that only a small number of people are working at HD+ resolutions due to hardware limitations, so I’m cheering them on.
If you want to go even higher in resolution, the next stable combinations are 1792×1008 for 16:9 and 1792×1344 for 4:3. At this resolution range, decoding after generation takes a noticeable amount of time, so it’s a good idea to increase the tile size to speed things up. In my case, I use the following settings in the WanVideoDecode node: tilex=640, tiley=640, stridex=512, stridey=512. For even higher resolutions, the options are 1920×1088 for 16:9 and 1920×1440 for 4:3. Although 1920×1080 is available as a preset in VideoConfig, it will likely result in an error. Please use custom_width and custom_height instead. At the time I released the workflow, I couldn’t properly test FHD tasks because they would immediately hit OOM.
@UncleHooru It's great to see your response! Thank you for the workflow. It is truly excellent—it almost never requires repeated generations, and I basically get great results on the very first try!
My ComfyUI skills are limited, so I look forward to seeing any future improvements you might make to the workflow to make the video generation even more stable. Thank you so much again!
Why not use V2 model, WanVideo Context Options node, why choose static standard mode, it seems that there is a slight flicker
You're lucky, I get light bulbs, light bulbs disappear, sometimes a lamp appears and then turns into something else. It's a shame, the person's quality was more than okay. But as for inconsistency...
@drak0n I tested it, and it didn't work very well, and neither did the sampler, and euler/beta didn't seem to work very well, and the same test clips, and the effect was far from the author's.
Speaking of v2 model, are you sure it is significantly better than v1 ? I never tried v2, I have v1 gguf Q8_0 from Kijai . I read that Kijai says something regarding that v2 difference as being better VRAM handling but quality might be worse than v1, You can read his comment here:
https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/discussions/27
@yTx I tried v2 right after it was released, but it didn’t work at the time. The rest of the setup has been stable on my system, so I’ve kept using it.
@drak0n If you feel like the model isn’t staying consistent, it’s usually because the target image you provided leaves room for the AI(WanAnimate) to decide what it thinks belongs there. For example, if there’s a strong shadow at the character’s feet, the model may assume there’s a ceiling light and try to generate one. Even a 1-pixel line near the edge can be mistaken for a wall.
Once you get used to it, you’ll start spotting these risky details in the target image before they cause trouble. Until then, it’s a good idea to practice with dark or solid-color backgrounds.
@hazzoom82659 I'll check v2 again.
@UncleHooru Thank you for your reply. I have tested countless "animate" workflows, and none of them generated elements that were not in the image (the background remains consistent). These inconsistencies can also be seen in the examples you provided. The image I used was of the highest quality, as was the video clip. This is the first time I've seen something like this, and believe me, I've tested many workflows. I like your workflow because it gives me consistency in the person's face, but when it comes to the consistency of the elements in the background, there is a big problem.
@UncleHooru euler/beta It seems to blur the background.
@drak0n Any parts that aren't visible in the initial image have to be generated by the model. Or does your model support multi-image input?
@yTx And if you ever find a scheduler and setup that make the background insanely sharp, let me know here.
@UncleHooru Don't get me wrong. I'm not criticizing you or your work. I can only imagine how much work and effort it takes to build such a workflow. Not to mention the time you spent accumulating all the knowledge necessary to build it. This is the first time I've seen these fluctuations in the background. I have tested numerous workflows, many of which offered poor quality but remained consistent in terms of background. That's why I'm trying to give you as much feedback as possible, so that you have a direction to try to find a solution to this problem. If it were general, I would have noticed it in other workflows as well, which was not the case.
Your workflow offers quality and consistency in terms of the person (something rarely found in other "animate" workflows) but encounters minor inconsistencies in the background. I can see with my own eyes that you are NOT a beginner. You are already building custom nodes, which requires a certain level of knowledge. I am the beginner, you are the expert. I congratulate you on that. There must be a solution to prevent such fluctuations. Now, don't be angry with me, I'm just trying to help you with a few impressions.
I've been waiting for this workflow for almost a month! I'm so excited!
I hope it works out. Good luck!
Does anyone know which node should be used instead of the "video_config_animate" node?
You can replace it by using a float node for the fps and int nodes for the other values. You can also simply set the values individually in each node without using them.
Thanks for sharing your workflow! I don't even generate videos locally, I still give you likes coz your videos are so good.
Thanks! Maybe this is a good chance for you to give it a try?
Hello, I’m trying to run WanVideoSampler in ComfyUI, but I’m getting a critical DLL loading error.
ImportError: DLL load failed while importing cuda_utils: The specified module could not be found.
🔻 My Environment
ComfyUI Distribution: StabilityMatrix (mapo build)
Python: 3.12.11
PyTorch: 2.8.0 + cu129
TorchVision: 0.24.1 + cu129
GPU: RTX 5090
CUDA / Driver: 12.9 (Driver 577.00)
OS: Windows 11 64-bit
Additional notes:
SageAttention installs but fails to load due to Triton not being found.
WanVideoWrapper loads, but WanVideoSampler fails immediately with the DLL error above.
Triton cannot be installed because PyTorch 2.8 does not provide a Windows-compatible Triton build.
🔻 What I’ve Tried
Reinstalled StabilityMatrix venv
Installed all missing Python dependencies (onnx, onnxruntime, opencv-python, scikit-image, mediapipe, protobuf, numpy, pillow)
Cleared pip cache
Reinstalled SageAttention
Updated GPU driver
Launched with TORCHDYNAMO_VERBOSE=1
The issue still occurs.
🔻 My Question
Could you please confirm the following?
Is WanVideoSampler compatible with Python 3.12 + PyTorch 2.8 + CUDA 12.9 (Windows)?
Is this DLL (cuda_utils) missing or unsupported in the current release?
What exact Python + PyTorch + CUDA versions are officially required?
Do I need to run a separate ComfyUI environment with Python 3.11 + PyTorch 2.5.1 (cu121) + Triton 3.0.0.post2 for WanVideoSampler to work?
Thank you for your help.
In my environment, it works with both Python 3.13 and 3.12.
I'm also using PyTorch 2.8 + CUDA 12.9 (Windows) just like you, and my GPU is also from the 50XX series.
As for installing Triton, if you follow the steps here
https://github.com/woct0rdho/triton-windows?tab=readme-ov-file#installation
you should be able to install it without any issues, and I believe there are plenty of explanations available as well.
The procedure is simply installing Triton via pip, and then—as stated in the manual—adding the two required folders to the python(_embedded) directory.
I'm using a portable version of ComfyUI, so that part is different, but you should definitely be able to get it working.
Thank you for your good work! Great workflow! Вut I still don't understand where to place "video_config_animate.py". I'm getting an error with the "VideoConfigAnimate" node? Thank you!
The file should be placed in /custom_nodes
Please refer to this scene in the tutorial video: https://youtu.be/DlJJs2roZco?t=204
I tested it,"Phantom_Wan_14B_FusionX_LoRA" will make the character over-sharpened, the actual experience is not too good, too clear, making the character feel fake.
please provide links to the onnx files needed for motion capture workflow! musch appreciated! great workflow!
I’m glad to see that it’s working well for you now.
lol yeah just looked for the onnx files on huggingface. Great workflow nice and clean/easy to tweak
@CharlesBarruls Hello, I see you've tested this workflow extensively. Have you tried any filters to improve the video quality? For greater realism?
@drak0n ive just adjusted the weights of the fusionx lora and the face and motion adherence. best advice would be starting with high fidelity source images and animation captures (videos you sample the motion from) best of luck. This is an excellent workflow imo.
@CharlesBarruls I also like this workflow, but I can't stabilize the background. That's the big problem. Otherwise, the workflow is one of the best I've used. But when, for example, a wall socket appears and then disappears, or a lamp appears and then disappears, it's not quite right. The other workflows I tested don't have this problem. I'm still looking for ways to reduce these inconsistencies. At the moment, I can't really find any.
@drak0n forgot to mention i also use the florence model captioner that provides the prompt a caption of the input image to reduce inconsistencies.
@CharlesBarruls Interesting approach. It may reduce some inconsistencies, but they are still there. This is strange. I would understand if it happened in general, but it seems to be specific to this workflow. While the other "animate" workflows have problems with the movement and consistency of the person's face, it seems that here they are solved, but another problem has arisen with the background.
@drak0n if you can please send me a .json of the workflow that is producing the inconsistent character motion but the stable background to compare.
@drak0n Have you tried dpm++_sde? Compared to eular/beta, it converges noise faster, so it’s effective for stabilizing the background. In that case, try setting the step count to 4.
is there any way to use it in 3060 please help its need very high ram nearly 60 gb
60gb? How long is your video ?
It may help to use a lightweight GGUF model to reduce memory pressure when offloaded.
https://huggingface.co/QuantStack/Wan2.2-Animate-14B-GGUF/tree/main
Can you create the workflow in Seaart? You can make videos for free using Comfy UI there.
Do you NEED 100GB of ram or does that just improve the efficiency?
Another question: I have an rtx 5080, but the generation cancels because I run out of memory. I followed the video 1-1, so I have no idea why. Might you have any ideas?
It doesn’t improve the efficiency. It just allows you to create longer videos. If you’re running into OOM crashes, instead of generating the full length, you could try testing with around 150 frames first. You can set the partial generation length using frame_load_cap in the VideoConfig (Animate) node. It’s also important to test under lighter conditions, such as lowering the resolution.
@UncleHooru lol son of a bitch I totally reinstalled comfyui cause I thought I fucked up myshit
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 68, 21, 80, 45] to have 36 channels, but got 68 channels instead
Please help)
I feel like I’ve seen this error during image generation before. Trying FP16 versions of the text encoder or VAE might help, or updating ComfyUI could also fix it?
I used fp16 bf16 fp32 vae, unfortunately it didn't help
@Vlad74477447 I had that error when I started using comfyui too. I think that happens when you mix vae, umt and diffusion model/checkpoint types. Since I use all of the same type I never got that error again. For example umt (text encoder) fp8 diffusion model fp8 and vae for fp8 (though not sure about vae, because many vaes are broken anyway, for example I have always to use vae wan21, even for wan22. Because vae wan22 never works in comfy ui for me.)
@bluenightlagoon thanks friend
Please can you help me understand what this error means: "Exception ignored in: <function ProactorBasePipeTransport._del__ at 0x000001DE0617C7C0>
Traceback (most recent call last):
File "asyncio\proactor_events.py", line 116, in del
File "asyncio\proactor_events.py", line 80, in repr
File "asyncio\windows_utils.py", line 102, in fileno
ValueError: I/O operation on closed pipe". Thank you in advance.
It seems to be a message that appears after the process ends, especially when an OOM error occurred just before it.
This workflow is so powerful, thank you！！！
Truly appreciate you sharing this! Do you have a workflow that you use to create your great initial images?
hi can you help me where to install the videoconfiganimate custom nodes
i cant find it
it should be in the zip file with the workflows, its a python file
Shouldnt the example files and workflow do fine with a 5090? I keep getting CUDA OOM issues
你好，经过我的尝试，请把wanvideo model loader里的attention mode 改为sdpa 。断开Torch Compile节点。wan....textencode...节点的device改为gpu。wanvideo block swap第一行参数40改为10（仅针对5090的优化）请试一试！
Hello, after my attempts, please change the attention mode in the wanvideo model loader to sdpa. Disconnect the Torch Compile node. Change the device of the wan....textencode... node to gpu. Change the first parameter of wanvideo block swap from 40 to 10 (optimization only for 5090). Please try it!
Hi, I really like this workflow! Would it be possible to make a version that replaces the character in the reference video with the one from the reference image? Most of the ones I've found are awful at handling more than 16fps and are really short by default without a lot of extensions