How to Make a Dance Video in ComfyUI: Current Native Video Workflows

The easiest way to make a dance-style AI video in current ComfyUI is to start from a native video workflow and add control only when you actually need it. Older tutorials often begin with AnimateDiff plus several custom-node packs. That can still be useful, but ComfyUI now ships official native workflows for modern video models such as Wan2.2.

What should you use in ComfyUI now?

ComfyUI's official Wan2.2 documentation currently includes native workflows for:

• a 5B hybrid model that supports text-to-video and image-to-video;
• 14B text-to-video;
• 14B image-to-video;
• first-and-last-frame video generation.

That is a better starting point than copying a long custom-node graph from an old video and hoping every dependency still exists. See ComfyUI's official Wan2.2 workflow guide.

Choose the kind of dance video before choosing nodes

“Make this person dance” can mean several different tasks:

Goal	Good starting approach
Create a new dancing shot from an image	Image-to-video with a strong motion prompt
Preserve a specific person's appearance	Image-to-video plus reference/identity control if needed
Follow a particular choreography	Motion/pose-reference workflow rather than prompt-only generation
Control the first and final pose	First-last-frame workflow

If you only want a believable dancing clip, do not begin with the most complicated motion-transfer graph. Generate a short image-to-video sample first. Add pose control when the exact movement matters.

Step 1: Update ComfyUI and load an official template

ComfyUI's documentation notes that its current video templates depend on recent core nodes. If a workflow template is missing or a core node cannot be found, an outdated ComfyUI install may be the reason.

For Wan2.2, open the workflow template library and search for the appropriate Wan2.2 template rather than assembling the graph from memory. The official guide also provides downloadable workflows.

Step 2: Pick a model that fits the machine

Model size determines much of the hardware requirement. ComfyUI's official Wan2.2 guide says the 5B hybrid version is designed to fit well on 8 GB VRAM with native offloading, while the 14B workflows are substantially heavier.

That does not mean every 8 GB system will produce the same speed. Resolution, frame count, text encoder, offloading, other software, and GPU architecture all affect memory pressure and generation time.

For iteration, start with the smaller workflow or lower resolution. Prove the motion and framing before spending more compute on a final render.

Step 3: Prepare a source image that can move

For image-to-video, the source image matters more than most node tweaks. A useful dance source usually has:

• the full body or enough limbs visible for the intended movement;
• clear separation between the subject and background;
• a pose that does not already contain impossible overlaps;
• enough image quality that the model is not guessing facial and clothing details.

If the source crops the feet but the prompt asks for full-body footwork, the model has to invent information that is not there.

Step 4: Describe motion, not just appearance

A useful video prompt states what changes over time. “A woman in a red jacket” describes a frame. “A woman in a red jacket performs a two-step dance, shifts weight left to right, turns once, full body visible, fixed camera” describes motion.

Include camera behavior when it matters. If the camera should stay still, say so. If the shot should track the subject, describe that. Otherwise the model may use camera movement to satisfy the feeling of motion.

Step 5: Generate a short test before adding more control

Use the native workflow to generate a short clip. Review four things separately:

1. Motion: does the body move in the intended direction?
2. Identity: does the person remain recognizable?
3. Anatomy: do limbs and hands stay plausible enough for the shot?
4. Background/camera: is unwanted movement stealing attention?

Change one variable at a time. If you alter the model, prompt, seed, resolution, frame count, and reference strength together, you will not know which change fixed or broke the result.

When should you add pose or motion reference control?

Prompt-only image-to-video is appropriate when “a convincing dance” is enough. It is not the same as copying a specific choreography. For exact motion, use a workflow that derives pose or motion information from a reference video and conditions the generation on that structure.

This is also where complexity rises. Pose extraction can fail when limbs overlap, the dancer leaves the frame, clothing hides joints, or the reference contains fast motion blur. Expect to clean the reference rather than treating it as a perfect control signal.

What about AnimateDiff, ControlNet, and IPAdapter?

They are not “obsolete,” but they should no longer be presented as the one universal ComfyUI video stack. AnimateDiff workflows remain useful for some Stable Diffusion-based pipelines. ControlNet-style pose or depth guidance can add structure. Reference adapters can help with appearance consistency.

The decision should follow the model and workflow you are actually using. Modern native video models may already solve part of the problem that older multi-extension graphs were built to patch.

Keep identity and motion as separate problems

If the dance looks good but the face drifts, adding more motion control may not help. That is an identity-consistency problem. If the face is stable but the choreography is wrong, stronger reference identity will not fix the motion.

Debug one axis at a time: motion, identity, composition, or temporal stability.

Rights and consent still matter

If the source image or reference video depicts a real person, use material you have the right to use and consider whether the generated result could falsely imply that person performed an action they did not perform. Do not treat technical ability as permission.

A practical current workflow

1. Update ComfyUI.
2. Load an official Wan2.2 image-to-video template.
3. Start with a clean full-body source image.
4. Write a motion-specific prompt.
5. Generate a low-cost test.
6. Only add pose/reference controls for a problem the test actually reveals.
7. Increase resolution or model size after the motion is working.

ComfyUI changes quickly, so use the official native workflow documentation as the version-sensitive part of the tutorial and keep your own workflow focused on the creative decision you are trying to control.