IAMCCS Wan-Animate-2 Extends

Wan-Animate-2 is officially a short-clip node: you give it a reference image and a driving video and it transfers the motion. "Extends" is the wrapper that turns it into a long-form engine. IAMCCS_WanAnimate2Extends samples the native WanAnimate2ToVideo in 4n+1 chunks, passes the previous generated frame through continue_motion so identity and motion survive across chunk boundaries, advances the driving-video offset each chunk, trims the latent/image anchors to the valid frame count, and - new in this edition - can composite the generated character over a background plate instead of whatever the model generated.

How it works

The loop is the whole story. Each iteration conditions on the reference image and a slice of the raw driving video (Wan-Animate-2 famously does not need a DWPose render - the pose_video input tooltip says it plainly), samples one 4n+1 chunk, then feeds the last generated frame back as the motion continuation for the next. The driving offset advances, so a long source video plays through in order. It validates the boring stuff up front - width/height divisible by 16, pose_start_percent ≤ pose_end_percent, a reference image and driving frames present - and clamps target_frames to the driving video's actual length (0 means "use all of it").

Chunk-to-chunk continuity comes from the context-window machinery it delegates to (ContextWindowsManualNode from ComfyUI core): context_length_latents (21), context_overlap_latents (8), context_schedule and context_fuse_method (pyramid/relative/flat/overlap-linear). There's also enable_pose_cache with CPU/GPU and int8/int4 caching so pose embeddings aren't re-encoded per chunk - the thing that makes long runs tolerable.

The inputs that matter

• reference_image + pose_video - the identity anchor and the raw driving frames.
• chunk_length - 81 is the official Wan-Animate-2 default; it's normalized to 4n+1 regardless.
• target_frames - 0 = full driving video; anything else caps it.
• cfg (1.0 default), reference_image_strength, pose_strength, pose_start/end_percent.
• reference_background_mode - keep_reference_background or isolate_character (needs reference_character_mask).
• output_background_mode - native_generated, or composite over source_video, reference_image, or a custom_background (background_image/background_video + composite_mask, where white selects the generated character).

Outputs: images (the composited final), native_images (pre-composite), plus frame_count, source_frames, trimmed_frames, and chunk_plan so you can verify the math.

Where people get burned

context_overlap_latents must be smaller than context_length_latents or it raises - that's a real constraint of the context-window system, not a style choice. And the isolate_character path needs a white character mask, otherwise the masking silently conditions on the wrong thing. The trimmed_frames output is worth watching too: it reports the anchor trim after each chunk, which is where the 4n+1 shortfall shows up if something's off.

It's the backend that the Easy Shotboard shell (IAMCCS_V2VShotboardEasyWanAnimate2) and the IAMCCS_WanAnimate2ShotboardBridge drive, so if you're using those you may never touch this node directly. Install the pack once - ComfyUI Manager (search "IAMCCS") or git clone https://github.com/IAMCCS/IAMCCS-nodes.git - and remember this expects native Wan-Animate-2 model/VAE/CLIP-Vision loaders upstream.