WanVideo Add Pusa Noise

Pusa is a rank-512 LoRA kijai extracted and hosts on his WanVideo_comfy HuggingFace repo, and it's popular enough that other projects lean on it too - the SCAIL-2 Infinity workflow, for example, wires it in and reports "noticeably better" quality with it than without. But Pusa doesn't behave like a normal LoRA you just bolt on at a strength value. It expects the model to sample under a specific scheduler mode - flowmatch_pusa - and that mode needs its own noise handling on top. This node is that noise handling: it adds latent and timestep noise multipliers specifically for the Pusa scheduling path.

Why it needs its own node

Ordinary flow-matching sampling denoises along a fairly predictable path from pure noise to clean latent. Pusa's approach perturbs that path - it introduces extra, controllable noise at specific points during sampling, which is part of what lets it (and things built on it, like SCAIL-2's chunked long-video generation) hold up better across boundaries and edits than a stock schedule. That extra noise has to be injected correctly or the LoRA's behavior falls apart, and that's not something you can bolt on generically - it needs to modify your image embeds directly, in the shape the flowmatch_pusa scheduler expects to read back out.

The inputs and output that matter

embeds (WANVIDIMAGE_EMBEDS) is your existing image-embeds bundle - this node modifies it rather than replacing it. noise_multipliers (default 0, range 0–100) controls how much extra noise gets injected; at the default of zero this node is effectively a no-op, so if you've added it expecting an effect and see none, check this value first. noisy_steps (default -1, range −1 to 1000) sets how many sampling steps the extra noise applies to - -1 reads as "all of them" in this kind of schema, so dial it down if you only want the effect early or late in the schedule.

Output is a single image_embeds (WANVIDIMAGE_EMBEDS) - the modified embeds, which continue on to your sampler exactly like any other embeds object, just with Pusa's noise behavior baked in.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. You'll also need the Pusa LoRA itself - kijai's Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors, from his WanVideo_comfy HuggingFace repo - dropped into your Wan LoRA folder and loaded through the pack's own LoRA-select node, plus a sampler set to the flowmatch_pusa scheduler.

Common issues & troubleshooting

Adding this node without the Pusa LoRA loaded does nothing useful. The noise multipliers are calibrated for how Pusa's weights respond to perturbation - inject them into a stock Wan model and you're just adding noise for no benefit, likely just softer or noisier output.

Wrong scheduler, no effect (or a broken one). This node assumes your sampler is actually running the flowmatch_pusa path. If your sampler is set to a standard scheduler like unipc or euler, the noise multipliers aren't being consumed the way they're meant to be - match the scheduler to the LoRA, not the other way around.

Start conservative on noise_multipliers. Since the default is 0 and the max is 100, there's a wide range here with no guardrail against pushing it too far. Community usage of Pusa (via SCAIL-2's integration) treats it as a quality-improving addition on top of an already-working pipeline, not a knob to crank - nudge it up gradually and watch for the point where extra noise starts reading as instability rather than improvement.