CLIP Sculpt Text Encode (Nukun)

Somewhere between prompt text and the image there's a layer of token vectors, and a small but determined community of people like to edit that layer directly. The tool that made this popular was CLIP Vector Sculptor - an external package that let you push token embeddings toward or away from each other ("sculpt" the prompt in embedding space). It's also the kind of personal project that ages: unmaintained, finicky with newer ComfyUI, a workflow liability.

NukunCLIPSculptTextEncode is a from-scratch, SD1/SDXL-native replacement for that old vector sculptor, folded into a single text-encode node. Same idea, no external package to babysit.

How it works

It's a CLIPTextEncode-style node (clip + multiline text in, conditioning out, plus a report) with a sculpting pass on top. The flow: tokenize your text, skip special and precomputed-embedding tokens (things like <lora:...> handles that shouldn't be nudged), then for each eligible token find its top_k nearest neighbor vectors in the CLIP embedding space and push the token toward or away from that neighborhood. The README's mechanism notes matter here:

• The nearest-vector search runs in bounded chunks with a CPU fallback - it doesn't clone the whole embedding table into VRAM, so it won't OOM your card the way the old package could.
• The "mean of all tokens" normalization is computed as a streaming chunk reduction rather than materializing a full FP32 copy.

That chunked/CPU-fallback design is the real upgrade. The old sculptor was notorious for memory pressure; this one is written to survive.

The inputs that matter

• sculptor_intensity - how hard to push (default 0.5, up to 5).
• sculptor_method - forward, backward, maximum_absolute, add_minimum_absolute. Forward pushes toward the neighbors; backward pushes away; the other two are variations on magnitude-based pushing.
• token_normalization - seven options including none, mean, and the default mean... actually the default is mean, and mean of all tokens is the streaming-chunk option mentioned above.
• top_k - how many nearest vectors to use per token (default 64).

The defaults (forward, 0.5, mean) are mildly active by design - the author's other sculpting nodes follow the same "mildly on" philosophy so you compare against a plain encode before turning anything up.

Why you'd reach for it

If you never touched vector sculptor, you probably don't need this - plain CLIPTextEncode is fine for most work. But if you've got a workflow that leaned on sculpting to push concepts together or apart (separating two similar subjects, sharpening a concept that keeps bleeding), this is the maintained, dependency-free way to keep that trick alive. The old Vector_Sculptor_ComfyUI package isn't patched or touched - so old workflows still load, and new ones can migrate here.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

Restart ComfyUI or install via ComfyUI Manager ("Nukun").

The honest caveat

Sculpting is a rabbit hole. It's easy to spend an afternoon pushing top_k and intensity around chasing a marginal effect, and the results are model- and prompt-sensitive. Treat this node the way the pack does: keep the defaults mild, verify the report shows the sculpt ran (and on which device), and don't expect magic - expect a tool for people who already know what "push this token away from its neighbors" means for their image.