PainterI2V for KJ

If you've run Wan 2.2 with a 4-step lightx2v LoRA, you know the drill: the generation is fast, the quality holds up, and everything moves like it's underwater. It's one of the most complained-about quirks of speed-distilled Wan - people were posting workarounds within weeks of 2.2's release. PainterI2V for KJ is a single-node fix for exactly that. It's a drop-in replacement for Kijai's WanImageToVideo encode node, and its entire job is to make the motion actually move again.

What it actually does

The node was written by princepainter as a modified fork of Kijai's Wan 2.2 I2V encode node - the README credits KJ right at the top, and the whole pack is one file. The pitch: +15-50% motion amplitude, better camera-prompt obedience, no brightness shift, and no quality loss at 1.0 (the "off" setting). Community testing backs it up - the thread that put it on people's radar was literally titled "Wan 2.2 Lightx + PainterI2V_for_Kijai" with the verdict "much more motions and speed."

The trick is a latent-space hack, not a new model. After your start image is VAE-encoded, the node separates the first-frame latent from the rest, computes the per-frame difference (the latent-space "motion vector"), strips the mean out of that difference so luminance stays put, scales it by motion_amplitude, clamps to [-6, 6], and recombines. You're literally amplifying how far each later frame drifts from frame one - which is exactly what reads as speed. The mean-preservation step is what keeps it from washing out or flickering into a brightness mess.

The inputs that matter

• motion_amplitude - the whole point. 1.0 = stock behavior, 1.15 default = general use, 1.25–1.35 for running/jumping, up to 1.5 for extreme motion. Tune in 0.05 steps.
• start_image - your I2V reference, required (it errors if you wire an image without a vae).
• vae, clip_embeds - feed the same WANVAE and CLIP vision embeddings your KJ workflow already uses.
• width / height / num_frames - output video geometry (defaults 832×480, 81 frames), matching the sampler.

Everything else - noise_aug_strength, start/end_latent_strength, temporal_mask, tiled_vae, the Fun/FLF2V toggles - is advanced or workflow-specific. Leave them alone on your first pass. The single output, image_embeds (type WANVIDIMAGE_EMBEDS), plugs straight into WanVideoSampler.

Installing it

No dependencies, no model downloads, no requirements.txt - it reuses the Wan 2.2 models and VAE you already have. ComfyUI Manager: search PainterI2VforKJ. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/princepainter/ComfyUI-PainterI2VforKJ.git

Restart ComfyUI. Here's the gotcha the README is quiet about: you still need Kijai's ComfyUI-WanVideoWrapper installed, because the model loader, VAE loader, sampler, and decode nodes in any working workflow are all wrapper nodes. This pack only replaces the encode step - it doesn't replace the stack around it.

Troubleshooting

• Motion still slow? Nudge motion_amplitude up in 0.05 steps, and make sure your prompt describes the movement ("running forward quickly"), not a vague "moves around." The README is right that prompting is half the battle.
• Too fast or jittery? Back it off. Pushing toward the 2.0 max is where you invite wobble and ping-pong-style artifacts - 1.4 is the author's own practical ceiling.
• VAE error on load? You wired a start_image but no vae input. Connect a WANVAE loader.
• Quality changed at 1.0? It shouldn't have - at 1.0 the node is a passthrough. If your baseline differs, the workflow, not this node, is the difference.

It's a niche tool for one specific annoyance, but if you live in 4-step Wan I2V workflows, it's a genuinely useful one - and it costs you nothing to try.