Image To Prompt

Point it at an image and it tries to write the prompt that could have made it. That's the whole idea - reverse the usual direction. easy imageInterrogator is Easy-Use's take on the classic CLIP-interrogator: image in, text prompt out. Handy when you've got a reference you like and want a starting prompt to riff on, or you're trying to match a style and need words for it.

Set expectations, though: interrogators give you a description, not a magic recipe. The output is a decent scaffold - subject, some style words, a few tags - that you then edit into a real prompt. Treat it as a first draft.

How it works

It runs a captioning/CLIP model over the image to produce candidate descriptive text. The mode picks how hard it works: faster modes give a quick caption, slower ones search more term combinations for a closer match. Output comes back as a prompt string you can paste into a text encode.

The inputs and outputs that matter

• image - the picture to describe.
• mode - fast, classic, best, or negative. fast is a quick caption; best searches harder and takes noticeably longer; negative targets negative-prompt-style terms. Start on fast to see if it's useful, escalate to best when you want a closer read.
• use_lowvram (default true) - keep this on if you're tight on VRAM; it trades a little speed to avoid blowing up memory. Turn it off only if you've got headroom and want it faster.
• prompt (output, a list) - the generated text. Wire it into a CLIP Text Encode or a text node to edit.

How to install it

ComfyUI Manager: search ComfyUI-Easy-Use, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat / pip install -r requirements.txt, restart. Heads up on the first run: interrogation needs its captioning/CLIP model weights, and if they aren't present the node will download them on first use. That first run can stall for a while as it pulls the models - it's not frozen, it's downloading.

Common issues & troubleshooting

First run hangs for ages. That's the model download, not a crash. Give it time on the first invocation; subsequent runs use the cached weights and are much faster.

best mode is painfully slow. Expected. best searches far more term combinations than fast or classic. Use fast while you're deciding whether the output is even useful, and reserve best for when you actually need the closest match.

Output reads like tag soup. Interrogators lean toward listy, keyword-heavy descriptions. That's normal - it's raw material, not a finished prompt. Prune the junk, keep the useful subject/style words, and write around them.

VRAM errors. If it OOMs, make sure use_lowvram is on. On a modest card, running this alongside a big model already loaded can tip you over - free some memory (or run the interrogation as its own step) before loading the diffusion model.