多文本CLIP批量编码(BNK)

MultiTextEncodeAdvanced (多文本CLIP批量编码(BNK), "multi-text CLIP batch encode") does exactly what its display name advertises: it takes a list of prompts, CLIP-encodes them all in one node, and returns a single batched CONDITIONING. The "(BNK)" is the giveaway - it's this pack's take on BlenderNeko's famous advanced CLIP text encode, vendored into the lam pack (src/adv_encode.py), with the list-processing machinery wrapped around it.

The use case: you have a list of prompts (say from MultiTextConcatenate's LIST output, or a batch text loader) and you want each frame of a batch to sample with its own prompt. Normally that means one CLIP Text Encode per prompt, a pile of wires, and a headache when you add prompt number six. This node turns it into one wire: prompt list in, one conditioning batch out.

How it works

For every string in textList, it runs the advanced encoder with your settings and appends the result. pre_text and app_text are optional prefix/suffix applied to every entry - so if all prompts share "best quality," type it once in pre_text instead of repeating it in every list item. The individual conditions and pooled outputs are concatenated into one CONDITIONING (and one pooled tensor), which the sampler consumes as a batch: latent frame 0 gets prompt 0, frame 1 gets prompt 1, and so on.

The settings that matter

• token_normalization - none, mean, length, or length+mean. Controls how per-token weights are normalized. mean is the safest general default; length+mean compensates for prompt length and is worth trying on SDXL.
• weight_interpretation - comfy, A1111, compel, comfy++, down_weight. This changes how (word:1.2)-style weights are parsed. comfy is stock behavior; A1111 mimics Automatic1111's parser if you're porting prompts over; compel and comfy++ are the power-user weight schemes.
• clip - your CLIP model. pre_text / app_text - optional wrapper text.
• Output: CONDITIONING.

If textList is empty the node raises - it wants at least one prompt.

Install

In yanlang0123/ComfyUI_Lam - Manager search "ComfyUI_Lam", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yanlang0123/ComfyUI_Lam

restart. No extra models; the advanced-encode code is bundled in the pack. The README's install.bat requirements aren't needed for this node, though you'll be installing the pack anyway if you came for the batch/list machinery.

Gotchas

The batch-coupling is the thing to wrap your head around: the conditioning batch and the image batch must line up. If your latent batch has 4 frames and your prompt list has 3 entries, ComfyUI will error out when the sampler tries to match them. Keep the list length equal to the batch size. Also - this encodes every prompt on every run, so a 50-prompt list is 50 forward passes through CLIP before sampling even starts; that's inherent to the approach, not a bug. And if you were expecting per-item conditioning you could select between (regional conditioning), note this returns one merged batch, not a selector.