WASABI CLIP Text Encode

It's a drop-in CLIPTextEncode that turns the embedding up

The WASABI pack's whole pitch is that your prompt is understood, but not strongly enough. WASABI CLIP Text Encode is the front door for that idea: it looks and behaves like the stock text encoder you already use - same clip and text inputs, same CONDITIONING output - but it doesn't stop at encoding. It takes the token embeddings and scales and interpolates them in ways that are supposed to push prompt adherence, with the author's focus squarely on Wan 2.1/2.2. The README's slogan is "because your prompts deserve flavor," and this node is the flavored version of the plain encode.

The honest thing to know up front: everything in this pack is experimental. All four WASABI nodes live in the for_testing category, and the author only claims to have verified results on Wan 2.1/2.2 text encoders. Treat it as a promising work in progress, not a settled tool.

How it works (and the one catch worth knowing)

Peeking at the source, this node is a wrapper. It calls the stock CLIPTextEncode, then hands the result to the pack's WASABI Conditioning Modulate node with your scale and mode. Here's the catch: the reference embedding it interpolates toward is the node's own output. Source and reference are the same embedding, so the interpolation phase has nothing different to pull toward - which means the mode you'll actually feel the effect of when using this node standalone is scale (it multiplies token magnitudes, then blends the result 50% back toward the original by default). The slerp-family modes come into their own when you want to stretch between two genuinely different prompt intents - that's the job of WASABI Conditioning Modulate with a real reference_conditioning wired in.

The inputs and output that matter

• text - your prompt, multiline.
• scale - default 1.5. Above 1 amplifies token magnitudes, below 1 dulls them.
• mode - pick of nine interpolation modes, default bislerp. For standalone use, scale is the one that visibly changes output; bislerp is the author's default for moving-toward-and-blending-back.
• clip - your CLIP or text encoder model, same as any encode node.
• advanced_options and schedule_options - optional DICTs from the pack's sibling nodes, when you want the fancier interpolation or per-token curves.

The single output is a CONDITIONING that plugs into your sampler's positive (or negative) input exactly like the output of any other text-encode node.

Install

ComfyUI Manager is the easy route: Manager → Custom Nodes → search "WASABI" → Install, then restart. Or from the terminal:

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

then restart ComfyUI. There's no model download and no heavy dependency - the only declared requirement is torch, with ComfyUI 1.3.0+ (the pack targets Wan 2.1/2.2, so you'll also need a Wan checkpoint and its text encoder, but those are the model files you already fetch for Wan, not WASABI extras).

Don't confuse this pack with WAS Node Suite - same author (WASasquatch, a fixture of the ecosystem since Easy Diffusion and the old 200+ node suite), but a completely different, tiny, new pack. WAS Node Suite is retired and breaks on current ComfyUI builds; WASABI is the new conditioning project.

Troubleshooting

• "I turned scale up and the image barely changed." In the interpolation modes, scale acts as a delta between the source and an identical reference - nothing moves. Switch mode to scale for a straightforward amplification, or feed a separate reference through Conditioning Modulate.
• Output looks the same as the plain encoder. With defaults that's expected-ish: the interpolation phase is degenerate by design. scale mode is where this node earns its keep standalone.
• Expect rough edges. It's for_testing category, and verified only on Wan 2.1/2.2. If you're on another model family and see nothing, that's the pack telling you it hasn't been tuned for your encoder yet.