If you have ever tried human image animation or pose transfer methods, you probably know the biggest headache. The identity of the person in the first frame keeps drifting as the video plays. Most existing methods follow a Reference-to-Video (R2V) approach, where they try to bind a static image to a moving pose. Sounds good in theory, but in real-world scenarios where the reference image and driving video often don't align perfectly in space or time the system panics. You get weird visual artifacts, abrupt transitions. That's the identity drift problem, and it has been extremely difficult to solve. SteadyDancer fix these problems.
 |
| Steady Dancer Working Pipeline (Ref- Official Page) |
Instead of forcing an image to follow a pose, it flips the paradigm and animate from the image itself. This Image-to-Video (I2V) approach ensures the most important thing is that your first frame stays intact. No more drift, no more mismatched poses, no more that doesn't look like the same person moments. The idea is simple but powerful that start from the reference state and evolve the motion from it. They defined clearly in their research paper. For in depth understanding you can go through into it.
Researchers behind SteadyDancer looked deeply into why identity drift happens. They identified two major culprits:-
-Spatial-structural inconsistencies-The reference image and the driving pose/video often donot share the same structure.
-Temporal start-gaps- The motion in the driving video rarely begins in sync with the reference image’s pose.
In R2V systems, these mismatches break the animation. But with the I2V paradigm, the model is trained to align motion to the image, not force the image onto mismatched motion. On top of that, SteadyDancer integrates three innovations:
-A Condition Reconciliation Mechanism that balances motion control and visual fidelity.
-Synergistic Pose Modulation Modules that adapt the driving pose so it fits the reference image.
-A Staged Decoupled-Objective Training Pipeline that optimizes motion, appearance, and coherence step by step.
Installation
1. First, install install ComfyUI if not yet. If already installed, update it from the Manager by selecting Update All option.
2. Make sure you have Kijai's custom node Wan Video wrapper installed. If already have then update the custom nodes from the Manager.
3. Download Steady Dancer models from Kijai's hugging face repository. Choose the one that suits your system resources:
(a) Steady Dancer FP8 (Wan21_SteadyDancer_fp8_e4m3fn_scaled_KJ.safetensors)
, for 12 to 16 GB VRAM.
(b) Steady Dancer FP16 (Wan21_I2V_SteadyDancer_fp16.safetensors), for higher VRAM 24 GB or more with better output. Save it inside your ComfyUI/models/diffusion_models folder.
4. Restart and Refresh ComfyUI.
Workflow
1. Get the workflow (wanvideo_SteadyDancer_example_01.json) inside your ComfyUI/custom_nodes/ComfyUI-WanVideoWrapper/example_workflows folder.
2. Drag and drop into ComfyUI. If you get missing red error nodes, just install them from Manager by selecting Install missing nodes option. The workflow is based on Wan 2.1 I2V framework. So all the basic models(wan 2.1 I2V model, wan2.1vae,
umt5-xxl etc ) will be same. Load them all into its relative node.
3. Execute your workflow by setting up the nodes:
(a) Load your image into Load image node. Then Load your reference video into load video node.
(b) Load Steady dancer model (Fp16 or FP8) into WanVideo Lora Select node.
(c) Load wan 2.1 Model into model loader node.
(d) Load wan 2.1vae, text encoders intotheir respective nodes.
(e) Add your detailed positive and negative prompts into prompt box.
(f)
Hit run to execute the workflow.
You can use block swapping(Value around 30-40) to run the model efficiently that will avoid any OOM (out of memory)errors.
KSampler Settings:
Scheduler-Dpm++
Steps-6
Shift-5
What we experienced is more longer video lose the face id consistency and sometimes it generates morphed videos.
