Chat with Image

Chat with Image is the pack's multimodal node: it takes an IMAGE from your graph, sends it to a vision-capable LLM, and returns the model's read on it as a string. Generate a character, dump it in, and get back a detailed description - which you can then feed right back into a prompt-generation node. It's the loop-closer for "generate, inspect, refine," and it's the node that makes an LLM feel less like a text box and more like a second pair of eyes on your renders.

How it works

Mechanically it's the Chat node with one extra step: the image tensor gets converted and attached to the request as an image part, alongside your prompt and optional system. Mustache rendering still runs on both text fields, so {{var}} templating works here too. The default prompt is "Describe this image." - which is a fine starting point, but the node really shines when you give it a specific instruction: "list every tag you'd use for this on Danbooru", "describe the lighting and camera angle", "what's inconsistent about this composition?"

Inputs that matter

• config - required, from API Config. Your model has to be a vision-capable one (most current flagship models are; check your provider).
• image - required, any IMAGE. From Load Image, from a VAEDecode, from a batch.
• prompt / system - what to ask, plus an optional system persona.
• vars, temperature, max_tokens - same set as Chat.

One output, text, the model's reply. Wire it to Markdown Preview to read it, or into a prompt-generation loop as the input to a Chat node that turns the description into a refined prompt.

The workflow it unlocks

The classic use is description → regeneration. Render, Chat with Image describes it, then a second Chat node takes that description plus your negative feedback ("the hands are wrong, the lighting is flat") and writes a corrected prompt. Run the loop and you've got a crude but genuinely functional self-critique loop. The pack's docs lean heavily on it as the eyes for the Anima workflow, but it's model-agnostic - any vision-capable model on your API Config works.

Installing and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/Moeblack/ComfyUI-SimpleChat

Restart, or use Manager and search "ComfyUI-SimpleChat". Only dep: aiohttp. No weights to download; the "model" is the remote vision API.

The gotchas are the same family as Chat: your API key lives in the workflow (scrub before sharing), and the model must actually support images or the request fails at the provider. One thing to watch specific to this node - the image is sent to a remote server, so if you're working on sensitive or private content, know that it's leaving your machine. And batch inputs: feed it a single image; a big batch in the image port can balloon the request.