Smart Painter Long Video

Naively chaining WAN I2V clips - render a segment, feed its last frame into the next - usually ends in drift: the character freezes, teleports to a new pose, or the camera lurches between segments. Smart Painter Long Video is the conditioning node that fixes that continuity problem, and it's the long-form sibling of SmartPainterFLF2V, built for clips up to 1000 frames instead of the usual 81.

It's the same execute logic as PainterVideo's PainterLongVideo (WanVideoWrapper pack) with one addition, and the mechanism is worth understanding because that's where the continuity magic lives. You feed it the tail of the previous segment via previous_video, optionally a start_image/end_image pair, and it builds the concat conditioning for the next segment: a gray canvas, the keyframes pasted in, VAE-encoded, and injected into your positive and negative conditioning as concat_latent_image plus concat_mask. The part that keeps long-form coherent is the motion reference - the last ~73 frames of the previous clip get encoded and passed along as reference_motion, and the previous clip's final frame becomes a reference_latent so the next segment starts where the last one ended. motion_frames (default 5) controls how many tail frames feed that motion reference, and motion_amplitude (default 1.15) nudges the model to keep moving instead of settling.

That's the trick most naive chains miss: the model sees where the motion was going, not just where the last frame froze. It's the difference between a character who keeps walking and one who resets to a fresh pose every segment.

The new bit versus the original PainterLongVideo is clip_vision_reference_image plus cv_reference_strength. It's an off-screen reference - a subject's identity, a jacket color, an object that never appears in the keyframes - merged with the existing clip_vision_output tokens along the token axis and fed through the model's img_emb cross-attention. That's the only non-keyframe image-guidance channel stock WAN 2.2 checkpoints expose without VACE or a ref_conv layer, so it's the one that works on plain WAN 2.2 I2V models. Just remember it's semantic, not pixel-faithful: great for "keep the jacket this color", weak for "reproduce this exact logo". Leave it unconnected and the node behaves exactly like the original.

Two gotchas worth knowing before you wire it up. First, initial_reference_image looks like the same feature but isn't: it's a VAE-encoded reference_latents entry that only has effect on checkpoints shipping a ref_conv layer (Fun Control, Animate, SCAIL, HuMo, VACE-merged). On stock WAN 2.2 it's silently ignored, so use the CLIP Vision slot instead for anything that should work everywhere. Second, the node raises an error if you connect none of previous_video, start_image, or end_image - it needs at least one keyframe source to build the canvas.

Outputs are positive, negative, and latent, which go straight into your WAN sampler. Feed the finished clip back into previous_video for the next segment and you've got an iteration loop that stays on-rails.

Install via ComfyUI Manager (search "Smart Helper Nodes"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/slvslvslv/ComfyUI-SmartHelperNodes

Then restart ComfyUI. No extra Python dependencies - the pack's requirements.txt is empty - and no model downloads beyond the WAN checkpoints you already have.