Wan Animate has been the king, but two new challengers have arrived: Steady Dancer and One-to-All Animation. Are they better? In this video, I test all three models and reveal a fatal flaw in Steady Dancer (the "looking down" head issue). But don't worry—I've built the "Ultimate Combo" workflow! By merging the pose recognition of One-to-All with the stability of Steady Dancer (and Wan), we can fix bad expressions and achieve the perfect AI animation results.
- - - - - - - - - - - - - - - - - - - - - - - - - -
✨ Register on RunningHub to get 1000 FREE RH Coins!
💰 Plus, claim 100 extra coins every single day just by logging in.
🚀 Try it Now:
1️⃣AI APP: Wan Animate+one to all:
https://www.runninghub.ai/ai-detail/1999835327776817154/?inviteCode=rh-v1291
2️⃣AI APP: SteadyDance+onetoall:
https://www.runninghub.ai/ai-detail/1999819498398117889/?inviteCode=rh-v1291
🔽Free download workflow and models/Loras 🔽
Download link for SteadyDancer FP16: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/SteadyDancer
Download link for SteadyDancer FP8: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/SteadyDancer
Download link for OneToAllAnimation_fp16: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/OneToAllAnimation/Wan21-OneToAllAnimation_fp16.safetensors
Download link for OneToAllAnimation_fp8: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/OneToAllAnimation/Wan21-OneToAllAnimation_fp8_e4m3fn_scaled_KJ.safetensors
Description
Wan Animate has been the king, but two new challengers have arrived: Steady Dancer and One-to-All Animation. Are they better? In this video, I test all three models and reveal a fatal flaw in Steady Dancer (the "looking down" head issue). But don't worry—I've built the "Ultimate Combo" workflow! By merging the pose recognition of One-to-All with the stability of Steady Dancer (and Wan), we can fix bad expressions and achieve the perfect AI animation results.
- - - - - - - - - - - - - - - - - - - - - - - - - -
✨ Register on RunningHub to get 1000 FREE RH Coins!
💰 Plus, claim 100 extra coins every single day just by logging in.
🚀 Try it Now:
1️⃣AI APP: Wan Animate+one to all:
https://www.runninghub.ai/ai-detail/1999835327776817154/?inviteCode=rh-v1291
2️⃣AI APP: SteadyDance+onetoall:
https://www.runninghub.ai/ai-detail/1999819498398117889/?inviteCode=rh-v1291
🔽Free download workflow and models/Loras 🔽
Download link for SteadyDancer FP16: https://huggingface.co/Kijai/WanVideo_comfy/tree/main/SteadyDancer
Download link for SteadyDancer FP8: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/SteadyDancer
Download link for OneToAllAnimation_fp16: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/OneToAllAnimation/Wan21-OneToAllAnimation_fp16.safetensors
Download link for OneToAllAnimation_fp8: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/OneToAllAnimation/Wan21-OneToAllAnimation_fp8_e4m3fn_scaled_KJ.safetensors
FAQ
Comments (37)
It's pretty amazing thanks.
It is following the reference pretty well, but there is a problem also : The begining works fine then after a few seconds of length, it starts blending cut from the reference image
Significant camera movements may trigger this issue, as the model remains trained on WAN2.1. We look forward to improved versions in the future. :)
@BiuBoomFlow Ah I see, while Wananimate is supposed to be 2.2 ? Also I've tried a "forcing" version of VACE with wan2.2 somewhere before.. It was badly optimized but pretty good also
@hboxgames132 Perhaps the name is indeed wan2.2, but in terms of quality, I don't perceive any significant improvement.🤔Moreover, the workflow does not differentiate between high-noise and low-noise models during loading, so I believe this still falls under wan2.1.
@BiuBoomFlow that's true.. unless it's Wan2.2 5B or something.. anyway looking forward for more stable progress
This workflow is awesome, thanks a lot for sharing!
I just have one question:
Is it possible to use Wan 2.1 or Wan 2.2 LoRAs with this workflow?
The quality is great overall, but I’m struggling a bit with physical body motion, like natural movement of body parts when a character jumps (for example, chest/breast jiggle).
Would using Wan 2.1/2.2 LoRAs help with this, or is there another way to improve that kind of motion?
Thanks again!
I have one more question.
In other workflows that use only Animate, even if the reference image shows a person sitting, the generated video usually starts with the pose from the sample video.
However, in this workflow, if the reference image shows a sitting person, the character keeps the sitting pose and starts dancing from that position.
Is this because this workflow works in a fundamentally different way compared to the other Animate-based workflows?
This is based on WAN2.1, so during certain exaggerated movements, distortion or bone loss may indeed occur. You can utilise WAN2.1's LORA, though the weighting requires testing on your part. In my own usage, values starting from 0.3 begin to yield results. Presently, I've only experimented with incorporating character LORA to maintain facial stability, but I haven't tested it for the chest area.
@rmrioe Yes, if you've watched my YouTube videos, you'll have seen that this SteadyDance model retains the reference image as the first frame of the action, then transitions to the movements within the reference video.
@BiuBoomFlow Thank you for the response! The quality is the best compared to any similar workflow I’ve used before. I’ll try using it more extensively.
@rmrioeA new model has now been released; you may wish to take note of it. The GitHub page is this one.:https://github.com/zai-org/SCAIL-Pose,It offers greater stability than SteadyDance in terms of motion, though it does not match SteadyDance's image quality for extended footage. I shall update the SCAIL-Pose workflow later.
@BiuBoomFlow Thank you! One last question though — in the final result, the person keeps smiling.
The original video has a completely neutral expression. Is there any way to fix this?
I’ve already tried adding prompts like “expressionless” and “serious expression”, but it doesn’t seem to make any difference.
@rmrioe Yes, I've also noticed the issue with facial expressions. Currently, the prompt within this motion transfer workflow doesn't have a significant effect. While the "Pose Detection OneToAll Animation" node allows adjusting the facial recognition mode, my tests indicate that disabling facial recognition causes the head to perpetually look downward again. Therefore, this workflow isn't viable. I have another workflow, not yet uploaded, that uses Wan Animate + OneToAll. This workflow can be configured to disregard facial expressions.
@BiuBoomFlow That sounds really exciting!! I had a lot of questions, and thank you so much for your kind and detailed answers!
@rmrioe We welcome further discussion and suggestions. : )
Very nice workflow. Can you please update the workflow with lora support as well
You may connect your desired LORA models before or after the accelerated LORA node. Based on my testing, the WAN2.1 model effectively stabilises facial features. The weighting need not be excessively high; I achieved satisfactory results using 0.3–0.5. As for other LORA models, I have yet to conduct testing.
Thank you for the suggestion. The workflow works well with WAN2.1 lora's as well
@oo7m I'm delighted to have been of assistance. Incidentally, your work is rather tempting.😍
Отличный workflow
а можно как-то убрать переходы на длинных видео? А то выглядит как неудачная склейка нескольких видео, может в настройках что-то?
В узле для загрузки видео на втором этапе есть опция для настройки скорости загрузки кадров видео, которая контролирует длину сгенерированного видео. Кроме того, в столбце, следующем сразу за скоростью загрузки кадров, можно указать количество пропускаемых кадров. Вы можете рассчитать количество пропускаемых кадров, а затем сгенерировать видео с интервалом в 81 кадр, чтобы получить высококачественный результат.
@BiuBoomFlow а можно номера нод в которых исправления нужно внести?
@habbibulinaglafira454 Конечно. На втором этапе рабочего процесса, в узле № 165 «Загрузить видео (загрузить)», вы увидите общее количество загруженных кадров в шестой строке. В седьмой строке вы можете установить количество кадров, которые необходимо пропустить перед эталонным видео.
Good job on the workflow bro, donated you 40 yellow buzz. enjoy!!
Thank you. Should you have any questions regarding usage, feel free to leave a comment for discussion.
@BiuBoomFlow Surely bro. For now your workflow is working entirely fine. If I come up with something I’ll let you know.
If you’d like we can connect on discord or any other place as you like. I am a creator as well.
I also thank you a lot for fixing the face down problem with the normal steady dancer workflow. The pose detection of onetoall on your workflow was a great idea of yours. Helped me a lot.
@coolstrad Ha ha ha, I'm delighted to hear you've got it up and running smoothly. I look forward to seeing your finished work, and I've followed you too. I don't have a Discord group at the moment, but if you have one, I'd be happy to join you lot.😆
@BiuBoomFlow I do have a discord server of myself you can join here. I am the owner of the server. https://discord.gg/RnUHPzTkXH
@coolstrad I've joined, though I won't be using it very often.🤣I've seen so many lovely pictures.
@BiuBoomFlow oh lol. i got some crazy stuff there haha. but great to have you. we can hacve some random chat there. I am the admin. my name there is DesireForDopamine
Thank you so much for the workflow! I've been having some problems when running it:
When using SteadyDance as a diffusion model, everything works fine until the point of sampling, when I get a cuda error and it crashes. I am using a 5070ti with pytorch version 2.7.0 cuda 12.8 because its the only version that apparently its compatible with my gpu. But also, when using Oneforall as a diffusion model, I get the error message that it's a text2video model instead of image2video mode. Can you please advise me? Thanks so so much!
First issue: Should the CUDA error stem from GPU memory overload, consider reducing the video generation resolution or increasing the Block Swap value, which can be set as high as "40". For Python version-related queries, it is advisable to consult GPT or Gemini directly, as this may expedite resolving environment configuration issues.
Second issue: When employing OneToAll as the diffusion model, the accompanying accelerated LORA must indeed utilise the T2V LightX2V model.
Thank you so much for the kind answer. I copied the log to chatgpt and it suggested to set false the use_non_blocking setting in the wanvideo blockswap node. That seemed to work, but I have the feeling it's a bit random, because I have tried with different input videos and so far it has only worked with one (but with different input images). Maybe the issue lies with the blackwell architecture still not being mature enough? Regarding the second issue, i downloaded the T2V Lora but got the same error message, i dont know what i'm doing wrong. Thanks again!
@jaysakurajimada1509 Regarding the first issue, I haven't encountered it myself; it's been running quite stably for me. Perhaps you could send me the complete erroneous code so I might be able to make headway. For the second issue, you'll need to download the T2Vlora version wan2.1. Here's the link:https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Lightx2v
Full filename:lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors
You may select different versions of Rang, but you must ensure it is Wan2.1.
能不能让奶子晃得明显一点
你可以通过接入lora提升晃动的幅度
This is a very nice workflow. But one issue keeps arising and I am not sure if there is a workaround to this, as the performer moves, the model has to invent newly exposed parts of the background, and each new segment may imagine those details differently, causing visible instability (after every 6 seconds). For longer shots on home hardware, pure AI generation usually struggles to stay consistent, so a hybrid workflow with background separation and compositing is the more reliable approach (for background consistency). Can anyone point me to the right direction?