AI video generation has improved dramatically over the past few years. Diffusion models can now create realistic movements, cinematic visuals, and impressive short clips from simple prompts. But there's still one major limitation duration. Most video generation models are designed to produce clips that last only 5 to 15 seconds. Once you try generating longer videos, the quality starts to break down. Characters changing appearance midway through the video. Dance movements becoming repetitive.
Instead of treating long videos as one enormous generation task, Wan 2.2 Dancer (built upon Wan 2.2 I2V by Tongi Labs) introduces a hierarchical framework that separates long-term choreography planning from frame-level visual refinement.
Rather than simply making videos longer, Wan 2.2 Dancer focuses on making them coherent from beginning to end.
 |
| Wan 2.2 dancer model architecture |
The first stage plans choreography using the global musical context. Instead of deciding movements frame by frame, the model understands the overall rhythm, tempo, and emotional progression of the music before generating detailed motion. A second refinement stage then focuses on producing visually rich and realistic dance movements while preserving consistency.
This separation significantly reduces temporal drift, repetitive motions, and identity degradation. This separation significantly reduces temporal drift, repetitive motions, and identity degradation. You can get more detailed insight by accessing their research paper.
Songs come in different lengths and tempos. To handle this, the model introduces Dynamic Frame Rate Adaptation, which maps RoPE (Rotary Positional Embedding) representations to absolute time. This allows the system to generate keyframes that naturally adapt to different song durations while maintaining accurate synchronization throughout the performance.
To improve visual quality, the researchers combine- Optical flow-based loss and Motion-speed control. These techniques help preserve smooth motion even during rapid choreography, enabling stable 720p videos at 30 FPS across multiple dance styles.
Installation
1. Make sure you have ComfyUI installed. Older user should update Comfyui from the Manager by selecting Update ComfyUI option.
2. Download Wan2.2 Dancer fp8(global and local). :
(a) Wan2.2 dancer 14b global fp8 scaled (wan2.2_dancer_14b_global_fp8_scaled.safetensors)
(b) Wan2.2 dancer 14b local fp8 scaled (wan2.2_dancer_14b_local_fp8_scaled.safetensors)
and save both of them into ComfyUI/models/diffusion_models folder. These will require at least 24 GB VRAM. If you have lower VRAMs use different variants.
(c) Wan 2.2 dancer GGUF variants, It ranges from (Q3-Q6). Use this model variant if you have low Vrams. Download both (local and global) of any variants and save them into ComfyUI/models/unet folder.
3. Now, download Lightx2v_I2V_14B_480p_cfg_step_distill_rank64 (Lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors ) and save this into ComfyUI/models/loras folder.
4. Download text encoder (umt5_xxl_fp16.safetensors) and save into ComfyUI/models/text_encoders folder.
5. Download Clip vision (clip_vision_h.safetensors) and put this into ComfyUI/models/clip_vision folder.
6. Finally, download VAE (Wan2_1_VAE_bf16.safetensors) and put this into ComfyUI/models/vae folder.
7. Restart and Refresh ComfyUI.
Workflow
1. Download Wan 2.2 Dancer workflow(Wan2.2_Dancer.json) from our Hugging face repository. If using GGUF variant just use this workflow and replace Load diffusion model node with unet loader node and make sure you have already installed the ComfyUI-GGUF custom node by City96.
2. Drag and drop into ComfyUI. If you find missing nodes, just update comfyui and install missing nodes from the Manager.
3. Load models (wan 2.2 dancer global and local model, clip, text encoder, vae, loras etc ) into their relevant node.
4. Now, upload your reference image and audio(music) file in mp3 format.
5. Set the configuration-motion_amplitude, dance_style, audio_duration, final_duration.
The model supports five dance style: K-pop, Chinese Classical, Street type , Latin and Tap Dance.
6. Hit Run to start the generation.
