WASABI Conditioning Modulate

This is where WASABI actually does its thing

The other three nodes in the pack are plumbing. WASABI Conditioning Modulate is the engine. It takes any conditioning you already have - from a stock CLIPTextEncode, a Wan T5 encoder, anywhere - and runs token-level math on the embeddings: scaling them, interpolating them toward a reference, blending them back. The README's promise is better prompt adherence and controllability on Wan 2.1/2.2, and this node is where that promise is delivered or not. It's also the honest answer to "I installed WASABI, now what do I wire where?" - you wire this.

Everything is experimental (for_testing category), and the author only validates Wan 2.1/2.2, so keep expectations calibrated. But the mechanism is real and worth understanding, because it's the difference between "I flipped a slider" and "I know why the image changed."

How it works

Conditioning in ComfyUI is a per-token tensor - roughly [batch, tokens, dims] - plus a pooled vector for global intent. WASABI operates on that directly:

• scale multiplies each token's embedding by a scalar (or a per-token schedule). Above 1 amplifies token strength, below 1 attenuates.
• slerp / lerp move each token toward a reference embedding. Slerp rotates along the arc between two vectors, preserving angular relationships (semantic-preserving); lerp cuts straight across, punchier but less norm-stable.
• The bi* modes - bislerp, bilerp, binlerp, bihybrid, biangleclamp, bisbezier, biease_slerp, biease_lerp - do the move-toward-reference, then blend back toward the original embedding by blend_back. That "interpolate, then soften the result" is what makes them feel less aggressive than a raw move.

keep_magnitude (default on) re-imposes the original L2 norm after each interpolation phase, so you change direction without accidentally making the prompt louder or quieter. Where a pooled output exists, the same logic is applied to it, so global prompt intent gets the treatment too.

The inputs that matter

• conditioning - required. Your existing CONDITIONING, from any CLIP/T5 encoder.
• mode - the nine modes above, default bislerp. For a first run, keep bislerp and a reference; drop to scale when you just want amplification.
• scale - default 1.0; >1 amplifies token magnitudes, <1 reduces.
• blend_back - 0–1, how much to blend the result back toward the original.
• reference_conditioning - optional, but the important one. The target to slerp/lerp toward.
• t_ref - default 0.5, how far toward the reference when no schedule is set.
• keep_magnitude and eps - norm preservation and numerical stability; leave them alone until you have a reason.

The output is a single CONDITIONING that feeds your sampler's positive or negative input like any other.

The trap: interpolation modes need a reference

Anything except scale hard-errors with ValueError: reference_conditioning required for this mode if you leave the reference unwired. Beginners hit this instantly. Wire a second conditioning in, or switch to scale. And the reference has to line up with your input: token counts must match (or be T==1), or you get Token length mismatch. Use a reference encoded from a prompt of roughly the same length.

Install

Via ComfyUI Manager (search "WASABI") or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/wasabi

then restart. No model downloads, torch is the only dependency, ComfyUI 1.3.0+ required. Same author as the famous (now retired) WAS Node Suite - different pack, don't cross the wires.

Troubleshooting

• "It won't run - reference_conditioning required." Pick a mode that doesn't need one (scale), or actually wire one in.
• "Token length mismatch." Your reference conditioning was encoded from a very different-length prompt. Re-encode it at a similar length, or use a T==1 reference.
• "I turned scale up and it sounds the same." With keep_magnitude on, the direction changes but the overall strength is pinned. That's the feature. Turn it off if you want brute amplitude.