Edit Image

Where DeepGen_T2I0 starts from nothing but text, DeepGen_I2I0 starts from a picture. You feed it one reference image plus an instruction, and a cloud model edits the picture to match - restyle it, change the lighting, remove or add an object, shift the mood, keep the subject and redraw everything else. It's the image-edit node of the DeepGen pack, and the "0" is the pack's way of marking it as the current generation (the old spelling is DeepGen_I2I, parked in DeepGen/Deprecated).

It earns its keep in two kinds of workflows. First, character or subject consistency: take a render you like, hand it back as image_1, and ask for variations of the same character in a new scene - the reference does the identity work that prompting alone can't. Second, fixing output locally: a generation comes back with a mangled hand or a wrong logo, and instead of re-rolling the whole thing you send it through an edit model with a targeted instruction. Both are things local img2img can do too, of course, but here the edit happens in the cloud on models you can't run at home.

Which models

The dropdown pulls from the pack's model list, filtered to edit-capable models. The set shifts as the pack updates, but you'll typically see Grok Imagine Image and Pro, Hunyuan Image 3.0 Edit, Wan 2.6, FLUX.2 [max] Edit, GPT Image 1.5, Seedream 5.0 Fast Lite, and the two Nano Bananas. Different models have different edit personalities - Hunyuan's edit variant is purpose-built for structure-preserving edits, while the Nano Banana family is better at following detailed restyle instructions. Try two and keep the one that matches how you work.

The inputs that matter

• image_1 (IMAGE) - your reference. Wire in anything producing an image: ComfyUI's built-in LoadImage, this pack's DeepGen_LIMG, or the output of another generator.
• prompt - the instruction, in plain English. "Keep the woman and the pose, replace the background with a rainy Tokyo street" beats tag soup here.
• minimum_resolution (500/1K/2K/4K), aspect_ratio, output_format - same trio as the T2I node; the pack maps them onto what the chosen model supports.
• nb_results, seed_value, output_prefix, config_json - standard across the pack, same semantics as DeepGen_T2I0.

One mechanism detail worth knowing: your reference image isn't sent as a file reference - the node converts the tensor to a base64 PNG and attaches it to the request. The pack even tries to trace the source filename through your graph and names the attachment accordingly (your LoadImage's file becomes image_1___mypic.png), which shows up on DeepGen's side and helps you keep track of what went in.

Outputs and wiring

• IMAGE - the edited result, a normal tensor. Save it, preview it, or chain it into another DeepGen edit for iterative passes.
• output_prefix_and_model (STRING) - model alias, useful for output naming.
• total_credits_used (FLOAT) - wire into DeepGen_F2T0 to meter spend. Edits are billable calls like anything else here.

Install and first run

Same pack as all the DeepGen nodes:

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 drop your DeepGen API key into ComfyUI Settings → "DeepGen API Key" (saved to ComfyUI/user/deepgen/config.json). No model files to download - requests and opencv-python are the only deps.

Gotchas

If you get a "DeepGen API Key not found" error, the key never made it into config. If the model list looks short, the pack's models.csv is stale - git pull to refresh it and restart. And since edits upload your reference image to a server you don't control, don't feed it anything you'd mind leaving your machine. If you want the same idea but with more references feeding the result, that's what DeepGen_I2I3 (three images) and DeepGen_I2IX (ten) are for.