Painter I2V Extend

Every Wan workflow ends the same way: a great five-second clip, and then the panic about what comes next. Feed the last frame into a fresh I2V pass and the third or fourth clip has drifted - the lighting shifts, the character's skin starts glowing like they just ran a marathon. That drift is the whole problem of long video, and Painter I2V Extend is built specifically to kill it. Instead of chaining clip-to-clip, it takes your previous clip and continues it, overlapping frames so the handoff is invisible.

It's one node from LDNKS094's ComfyUI-Painter-I2V-AIO pack, which is a cleaned-up, unified rewrite of princepainter's scattered Painter nodes for Wan 2.1/2.2. Extend itself is the successor to PainterLongVideo. It doesn't sample - it's a conditioning node. You feed it your old clip, it returns the conditionings and latent you hand to a sampler, and the sampler (PainterSampler in the same pack) does the denoising.

How it works

Two modes, picked by one toggle. The default, svi_mode=False, is the standard continuity mode: it takes previous_video[-overlap_frames] as the start frame, uses the last frame of your clip as a "middle" anchor at position overlap_frames, and auto-calculates a middle strength so you don't get the gray artifacts that plague naive frame-concat tricks. The more frames you overlap, the smoother the join.

svi_mode=True switches to the SVI 2.0 Pro architecture: an anchor frame (your anchor_image, or previous_video[0] if you leave it empty) plus the last overlap_frames encoded as motion, padded with latents_mean - a zero-valued latent that SVI LoRAs specifically expect. This is the mode that gets you the 20–25 second continuous takes SVI is known for on Wan 2.2. Don't touch it unless you're actually running an SVI LoRA; the padding trick is meaningless without one.

Both modes share the two knobs that make Painter nodes famous. motion_amplitude (default 1.15) fixes the slow-motion problem of 4-step LoRAs like LightX2V by amplifying the difference between anchor and other frames - 1.1–1.2 for normal motion, 1.2–1.5 for running, jumping, action. And color_protect (on by default, leave it on) detects channels with mean drift after that amplification and gently corrects them, which is why your skin doesn't go green.

The inputs that matter

You only really set a few of these:

• previous_video - the clip you're continuing (required).
• overlap_frames - 4 for a standard join, 8 for a smoother transition. Standard mode only.
• motion_amplitude - motion kick from your speed LoRA.
• svi_mode - off unless you run SVI.

Then width, height, length, batch_size, positive, negative, vae behave like any Wan I2V setup. Optional end_image gives you first-last-frame anchoring (FLF2V), and clip_vision adds semantic guidance. Outputs: positive, negative (CONDITIONING) and latent - all three go straight into the PainterSampler node.

Installing it

ComfyUI Manager, search "Painter I2V AIO" (or the repo name). Or the old-fashioned way:

cd ComfyUI/custom_nodes
git clone https://github.com/LDNKS094/ComfyUI-Painter-I2V-AIO

Restart ComfyUI. No Python dependencies - it's pure in-tree code, so the real install burden is the model stack: a Wan 2.1/2.2 I2V checkpoint, its VAE, text encoder and CLIP vision, plus a 4-step LoRA if you want the motion fix. One gotcha: the README still references the old "PainterAIO" repo name in a few places - clone the ComfyUI-Painter-I2V-AIO URL above.

Where people get burned

Cranking motion_amplitude past ~1.5 turns motion into shaky, hyperactive messes - the node amplifies whatever motion exists, it doesn't invent clean physics. And this is a V3-API node, so if it's missing from your node list, update ComfyUI first. The classic failure this node exists to fix is treating it as a fresh I2V: without the overlap frames, you're back to clip-chaining and the sauna-skin drift returns.