Why do my results look soft or low-detail?

Wan 2.2 Fun Control uses a 5B-parameter model prioritized for compactness and experimentation. Improve perceived quality by keeping scenes simple, using strong and clean Canny edges, reducing clip length, increasing sampling steps moderately (e.g., 20–30), and lowering CFG (3–6) to reduce oversaturation and noise.

Can I use depth or pose instead of edges?

Yes, the model supports multiple conditions. This template wires Canny by default; to use depth or pose, add appropriate nodes (e.g., a depth estimator or a pose detector) and feed their outputs into the corresponding inputs of Wan22FunControlToVideo.

What frame rate and duration should I use?

The model was trained with clips at 24 FPS and up to around 121 frames. Matching 24 FPS and keeping clips short (2–5 seconds) generally yields more stable motion and fewer artifacts with this 5B model.

How do I get more consistent motion from a reference video?

Load your source via LoadVideo, route it through GetVideoComponents to preserve fps and frame order, and generate a clean edge map with Canny. Increase control strength in Wan22FunControlToVideo if available, and lower CFG to reduce hallucinations. Using a fixed seed helps with reproducibility.

Wan 2.2 5B Fun Control - ComfyUI Workflow

Wan 2.2 5B Fun Control is a compact, multi-condition video diffusion workflow that demonstrates how to steer motion and composition using control signals—most notably edges via the Canny node—while generating short clips from text, images, or reference video. The core of the graph is the Wan22FunControlToVideo node, which fuses your prompt (from CLIPTextEncode), optional start image or video (LoadImage, LoadVideo + GetVideoComponents), and control maps (Canny) into a temporally coherent latent sequence. A UNETLoader loads the wan2.2_fun_control_5B_bf16.safetensors diffusion model and pairs it with VAELoader (wan2.2_vae) and CLIPLoader; sampling is driven by KSampler using ModelSamplingSD3 to match Wan 2.2’s scheduler. VAEDecode converts latents to frames, CreateVideo packs frames to a clip, and SaveVideo writes the final file.

Technically, the workflow is organized into clear groups: Step1 - Load models, Step 2 - Start_image, Step3 - Prompt, and Step 3 - Control video and video preprocessing. Edge guidance is prepared by the Canny node (low/high thresholds) and passed to Wan22FunControlToVideo to anchor structure and motion. You can preview intermediate frames with PreviewImage, then finalize your result with CreateVideo and SaveVideo. While Wan 2.2 Fun Control supports multiple conditions (e.g., depth and pose), this template wires up edge control by default; you can extend it by adding depth/pose nodes and routing their outputs to Wan22FunControlToVideo. Note: at only 5B parameters, Wan 2.2 Fun Control trades fidelity for speed and accessibility—expect softer details and lean on strong control signals, simpler scenes, and shorter clips for best results.

FAQ

常見問題