CLIP Prompt Conditioning 🐼

What it is

This one's a fiddly, powerful CLIP conditioning tool, ported into this pack from Extraltodeus/Vector_Sculptor_ComfyUI - the README says so directly. If you've never touched Vector Sculptor, think of this as an alternate CLIP Text Encode with extra knobs for blending multiple prompt lines together and reshaping how strongly the encoded vector expresses each concept, beyond what standard (word:1.3) attention weighting gives you.

Worth saying up front: this is CLIP-era machinery. It only does anything meaningful on models that still use a CLIP text encoder - SD 1.5, SDXL, and the Illustrious/NoobAI/Pony lineage. If you're prompting a 2025-2026 LLM-encoded checkpoint (Z-Image, Flux 2 Klein, Anima), attention weighting and vector manipulation of this kind aren't part of how those models read a prompt, and reaching for this node there won't do what you expect.

How it works

You give it a CLIP model and a block of text - write more than one line and each line gets encoded and merged into a single conditioning, rather than being treated as one long string. The merge happens one of two ways: slerp (spherical interpolation, which moves along the surface of the unit hypersphere between the two conditioning vectors rather than just averaging their coordinates) or a plain average. On top of the merge, "sculpting" reshapes the resulting vector's magnitude along a chosen direction, which is the Vector Sculptor technique this node is built on - it's a way of pushing concepts harder or softer than standard attention-weight syntax allows.

Inputs and outputs

• clip - your CLIP model.
• text - multiline; write one prompt/concept per line if you want them merged.
• merge_conditioning_type - average (default) or slerp.
• merge_conditioning_strength - 0-1, default 0.5, how much weight the merge gives to each line.
• merge_conditioning_strength_custom - override the single strength value with a per-line custom string if you need finer control than one global number.
• sculptor_intensity - 0-10, default 1. How hard the sculpting effect pushes.
• sculptor_method - forward, backward (default), or maximum_absolute. Different directions/strategies for how the reshaping is applied.
• token_normalization - seven options (none, mean, set at 1, default × attention, mean × attention, set at attention, mean of all tokens), controlling how per-token attention weights get renormalized after encoding - the same category of knob as A1111/compel-style normalization settings.

Output: a single CONDITIONING, wired straight into your sampler like any other conditioning.

Installing it

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/zhongpei/Comfyui_image2prompt

Or search "Comfyui_image2prompt" in ComfyUI Manager. (The README's own snippet has a typo'd, hyphenated URL - use the underscored one above.) No model download needed beyond the CLIP checkpoint you're already using in your workflow; this node borrows your CLIP input rather than loading its own.

Common issues

Using it on the wrong architecture. This is the big one - if your checkpoint is LLM-encoded, none of this node's controls will move the needle, because prompt weighting syntax and vector-space reshaping are CLIP-specific mechanisms that don't exist on that prompting path. Check what your checkpoint is built on before reaching for this.

Too many knobs, no baseline. With seven token-normalization options and two merge strategies plus a sculpting pass on top, it's easy to turn three dials at once and lose track of what changed anything. Start with the defaults - average, strength 0.5, sculptor intensity 1, backward - get a working bake, then adjust one setting at a time.

Sculptor intensity above a few points can distort output the same way stacking attention weights too aggressively does on plain CLIP Text Encode - if results start looking artifacted rather than just "stronger," back the intensity down before touching anything else.