Generate Image (from 3 Images)

One reference image is good for keeping a subject recognizable. Three is how you tell a cloud model "this is definitely the same person/character, don't argue." DeepGen_I2I3 is the multi-reference image generation node in the DeepGen pack: feed it up to three images plus a prompt, and it produces a new image that has to reconcile all of them - the classic character-consistency move, done entirely in the cloud.

Why you'd use three references

This is the node for identity work. A single reference can drift when the model decides your prompt matters more than the picture. Three angles - a face-on shot, a profile, a full-body - give the model enough evidence that the thing it's drawing is that subject, not "a person who looks vaguely similar." It's the same trick the community uses with reference-heavy closed models, which are built for exactly this. If you're iterating on an OC, a character sheet, or a consistent product shot, this is the DeepGen node you'll reach for before the single-reference edit node.

The inputs

• image_1, image_2, image_3 (IMAGE) - your references, all optional. You can use fewer than three; the node doesn't care if some stay unconnected. Wire in LoadImage, DeepGen_LIMG, or whatever you have.
• prompt - what you want the result to do, on top of the references: "the woman from the references, standing in a sunlit garden." The model merges subject identity from the images with scene from the text.
• model - dropdown of multi-reference-capable models: Hunyuan Image 3.0 Edit, Grok Imagine Image, Wan 2.6, GPT Image 1.5, Seedream 5.0 Fast Lite, Nano Banana Pro, Nano Banana 2. They vary in how faithfully they hold identity across multiple references, so expect to trial two or three.
• minimum_resolution (500/1K/2K/4K), aspect_ratio, output_format - the usual trio, mapped onto the selected model's supported sizes.
• nb_results, seed_value, output_prefix, config_json - standard pack behavior; config_json is the escape hatch for extra API params.

As with the other image nodes, each reference tensor becomes a base64 PNG attachment on the request, and the pack tries to preserve source filenames by tracing your graph, so what you sent stays identifiable.

Outputs

• IMAGE - the generated image batch. Save it, preview it, or chain it into another DeepGen node.
• output_prefix_and_model (STRING) and total_credits_used (FLOAT) - the same bookkeeping pair every DeepGen generator emits.

Install

Same pack as the rest:

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" (stored at ComfyUI/user/deepgen/config.json). Only requests and opencv-python are installed - everything heavy is server-side.

Gotchas

Each generation spends credits and uploads your reference images to the cloud - don't send private work you're not comfortable leaving your machine. If identity still drifts with three references, that's the model's limit, not a settings bug; try a different model from the dropdown, because the reference-fidelity ranking changes between them. And note this node is current - there's no deprecated spelling for DeepGen_I2I3, unlike its single-reference cousin.