Stack_Wan22FunControlToVideo

Wan's "Fun Control" line is Alibaba's own official answer to the same problem VACE solves - driving video generation with a control video (depth map, pose skeleton, edge map, whatever your preprocessor produced) instead of relying on the prompt alone to describe motion. It's a real, working model family - people have used it for things like driving 3D-animated shots with AI-generated detail on top, one showcase running Wan 2.1 14B Fun Control at 720p to layer AI polish onto existing CGI. Stack_Wan22FunControlToVideo is this pack's config node for the Wan 2.2 version of that model.

How it fits together. This isn't a sampler - it's a settings bundle. You set your target width, height, and length (frame count; the default 81 matches the community's standard ~5-second clip at 16fps), optionally hand it a ref_image for identity/style anchoring, and a control_video - your preprocessed driving footage. The output, funControl22, feeds into a downstream Wan 2.2 sampler that actually runs the two-pass, high-noise/low-noise MoE generation Wan 2.2 uses under the hood.

What the two optional inputs actually do. control_video is doing the heavy lifting: it's what constrains motion and composition frame-by-frame, the way a depth or pose sequence would in an image ControlNet, just applied across time instead of to one frame. ref_image gives the model something to anchor appearance to, separate from whatever motion the control video is dictating - useful when you want the control video's movement but a different subject's look than whatever the control footage originally contained.

Installing it. No extra dependency specific to this node beyond the base pack install: git clone https://github.com/cardenluo/ComfyUI-Apt_Preset into custom_nodes, restart ComfyUI, or search "ComfyUI-Apt_Preset" in ComfyUI Manager. You will need the actual Wan 2.2 Fun Control checkpoint files downloaded separately (Alibaba's official releases, not something this pack bundles), plus a preprocessor upstream to turn your source footage into the depth/pose/whatever control video this node expects - this node has no built-in preprocessing, it just consumes an already-processed control video.

Common issues. Motion that doesn't track the control video at all almost always means the control video wasn't preprocessed into the format the model expects - feeding it raw footage instead of a depth or pose extraction is the most common version of this mistake. If output looks incoherent or the character keeps morphing, that's frequently a sign the control video's resolution or frame count doesn't cleanly match width/height/length - Wan's frame math is picky (video frame counts want to land on a 4n+1 pattern internally), so mismatches here tend to produce visibly broken results rather than a clean error. And because Fun Control is a genuinely separate model line from VACE, don't expect VACE-trained LoRAs or workflows to drop in cleanly - match your control video's preprocessing style (pose vs. depth vs. edge) to what the specific Fun Control checkpoint you downloaded was actually trained on, since the README for the underlying model usually specifies which control types it supports.