Review Images

The inverse of an image generator: DeepGen_I2T0 is a vision-language node that reads your images and writes text back. You feed it up to ten images plus an instruction, and a cloud VLM - Gemini 3 Pro, Claude Opus 4.6 / Sonnet 4.6, GPT-5.4, Gemini 3 Flash, Gemini 3.1 Flash Lite - looks at them and answers in a string you can wire anywhere in the graph.

Why you'd reach for it

Three jobs, in rough order of popularity. QC on a batch: generate four variations, send them all in, and ask "which of these best matches the prompt, and why?" - the model does the triage you'd otherwise do by eye. Captioning for re-use: ask for a detailed description of a render and feed the text back into an image or video generator as a prompt, which is the caption-to-seed workflow the KB's VLM section describes. Iterative editing direction: ask the model to critique a result and spit out a revised prompt, then feed that into an edit node. Because it takes up to ten images, it doubles as a mini batch-reviewer - a genuinely handy way to eyeball a whole grid without eyeballing a whole grid.

The inputs

• image_1 through image_10 (IMAGE) - all optional; wire in one or all ten. Each is converted to a base64 PNG attachment on the request, and the pack traces source filenames so DeepGen's side can tell them apart.
• prompt - the instruction that defines the job. Be explicit about what you want back: a JSON structure, a comma-separated caption, a one-paragraph critique. "Describe this image" is fine but wastes the multi-image budget; "compare images 1–4 and pick the sharpest" is what this node is for.
• model - the VLM dropdown. Gemini 3 Flash / Flash Lite are the budget reviewers; Pro and Claude are the better writers.
• seed_value, config_json - standard pack fields.

The outputs

• output (STRING) - the model's text. This is the payload; wire it to a prompt input, a display, or a text save.
• model (STRING) - which model answered, for logging.
• total_credits_used (FLOAT) - meter it with DeepGen_F2T0.

Note this node has no nb_results or output_prefix - it's a single text answer, not a batch generator.

Install

Same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/deepiksdev/ComfyUI-DeepGen-API.git
cd ComfyUI-DeepGen-API
pip install -r requirements.txt

Restart, then set your DeepGen API key in Settings → "DeepGen API Key" (ComfyUI/user/deepgen/config.json).

Gotchas

The same output-hygiene warning as the LLM node: if you ask for a caption, the model may pad it with chat scaffolding, and none of that belongs in a downstream prompt. Ask for bare output. Also remember every call sends your images off-box and spends credits - the ten-image review is cheap per image but it all adds up. And if you spot the old name DeepGen_I2T ("⚠️ [Deprecated] Review Images") in a legacy workflow, that's this node's pre-rename spelling; use DeepGen_I2T0 for new work.