CLIPtion Generate

The name is a lie. CLIPtion Generate doesn't call any API, needs no key, and doesn't load some big vision-language model on the side. It's a tiny captioning decoder that rides on the CLIP-L text encoder and CLIP-L vision encoder already sitting inside your checkpoint. If you're running SD, SDXL, SD3, or FLUX, you're already paying the VRAM cost of ViT-L/14 - CLIPtion just borrows it to write captions. That's the whole trick, and it's why this thing generates captions at roughly the speed of a blink.

You reach for it when you want a prompt in the workflow, not from a separate tool: auto-captioning a folder of images, feeding a generated image's description back into img2img or a text-to-image pass, or roughing out captions for a LoRA dataset before you clean them up by hand. The author is upfront about the trade-off - bigger dedicated captioners and VLMs will caption more accurately - and the community agrees. In the announcement thread people reported it faster than Florence and competitive with JoyCaption on simple concepts, but it clearly misses harder content and short-circuits on NSFW. Treat it as the fast convenience captioner, not the dataset gold standard.

Here's the mechanism. You feed a CLIPTION model from CLIPtionLoader plus an image. The image goes through the CLIP vision encoder, and then a small transformer decoder (six blocks, 768 hidden, 8 heads) generates tokens one at a time, up to CLIP's hard 77-token ceiling. Three knobs change how it writes. temperature (default 0.7) sets how random the sampling is - higher is more diverse, lower is more predictable. best_of (default 1, up to 64) samples that many candidate captions in parallel and picks the one whose text embedding has the highest CLIP similarity to the image, which is a neat self-correcting trick: you don't just hope the sampling was good, you grade the candidates. ramble forces the decoder to produce the full 77 tokens instead of stopping at its end-of-sentence token.

The inputs that matter, in order of how often you'll touch them:

• image - any image or batch; the node captions each frame and returns one string per image.
• seed - required, because this is sampling. Different seeds give different captions.
• best_of - the one you'll actually turn up. Higher = better captions, more compute.

The output is a STRING list (one caption per input image). Wire it into a Show Text / Preview Text node to read it - the pack's example workflows use pythongosssss's ComfyUI-Custom-Scripts for that - or straight into a prompt encode node if you're chaining caption-to-generation.

Install is the usual: ComfyUI Manager, search "cliption", click install, restart. Manual route is cd ComfyUI/custom_nodes && git clone https://github.com/pharmapsychotic/comfy-cliption.git && pip install -r comfy-cliption/requirements.txt, then restart. Dependencies are refreshingly light - just huggingface-hub, safetensors, and transformers.

Where people get burned: the loader throws "Must use model which includes CLIP-L" if your checkpoint lacks the CLIP-L text encoder (any mainstream SD-family model has it, so this is usually a wrong-model pick, not a real problem). And you need a CLIP vision encoder wired in - install openai/clip-vit-large-patch14 via Manager > Model Manager > search "clip vision large". If captions feel short, that's the 77-token CLIP wall doing its job; flip ramble if you want the full-length ramble. If they feel wrong, that's not a bug - this is a 100MB captioner doing its best. For the money, it's remarkably good.