GPT Image Generate

If the general ChatImageBridge node is your Swiss-army endpoint client, GPTImage is the one tuned specifically for the GPT Image / Nano Banana family - gpt-image-2-vip, gpt-image-2, nano-banana, nano-banana-2, and nano-banana-pro. These are the closed Google/OpenAI image models the KB's closed-source-models.md catalogs: Nano Banana Pro is Gemini 3 Pro Image with 4K native output and "thinking mode", Nano Banana 2 is the 3.1 Flash speed-quality hybrid, and GPT Image is OpenAI's natively-multimodal image model. You're not running any of that locally - this node is a client for a relay that fronts them, which is the whole point. The heavy lifting happens in someone else's data center; your GPU stays cool.

How it works

Where the generic node uses non-streaming chat completions, this one streams the response ("stream": true). The README says that's deliberate: image generation can take a while, and a streaming request is far less likely to trip Cloudflare or an upstream gateway timeout than sitting on one long non-streaming call. When the provider returns an image as a direct URL in the stream, the node downloads it and converts it to a tensor; it also handles b64_json and markdown references.

The part worth understanding is the size math, because it's not just passed through. When both resolution and aspect_ratio are explicit, the node looks up a concrete pixel size and sends it as the size field:

4K + 16:9 -> 3840x2160
2K + 1:1  -> 2048x2048
1K + 9:16 -> 720x1280

If either is auto, it doesn't error out - it just skips size and writes your intent into the message as a hint ("Target image resolution: 4K."), letting the provider interpret it semantically.

Inputs and outputs

The essentials are api_key, base_url, model (a dropdown with the five models above), and prompt. Then resolution (default 1K), aspect_ratio (default 1:1), and timeout_seconds (default 600, 60–3600 - generous, because streaming long generations is the whole design). image_1 and image_2 are optional and get attached as reference images when present. Output is a single image tensor, straight to Save Image.

The trap: resolution limits

Not every model takes every resolution. The node enforces this per model:

| Model | Allowed | | --- | --- | | gpt-image-2-vip | 1K, 2K, 4K | | gpt-image-2 | 1K only | | nano-banana | 1K only | | nano-banana-2 | 1K, 2K, 4K | | nano-banana-pro | 1K, 2K, 4K |

Pick 2K on plain gpt-image-2 and the node throws on purpose rather than fire a doomed request and burn your credits. The error message tells you what the model supports. People hit this constantly because the default resolution is 1K and then they bump it - read the error instead of fighting it.

Install

Part of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Marigold1122/Chat-Image-Bridge-Node.git ComfyUI-ChatImageBridge
pip install -r ComfyUI-ChatImageBridge/requirements.txt

Restart, then api -> Chat Image Bridge -> GPT Image Generate (or ComfyUI Manager, search "Chat-Image-Bridge-Node"). Just requests for a dependency.

Troubleshooting

• SSL errors behind a proxy - the streaming code has a real-world quirk worth knowing: if the first request hits an SSL error, it retries once with environment proxy settings disabled, because corporate proxies frequently break these relays.
• "GPT Image stream did not contain image data" - the provider streamed text but no image, usually a refusal or a non-image response. The error shows a preview of what came back.
• Long generations - if the stream just hangs, check timeout_seconds; the default 600s is there because these models genuinely take minutes at 4K.
• A streaming provider must actually stream - a few relays advertise compatibility but return a single non-streaming blob, which this node also tolerates in practice. If yours fails consistently, the plain ChatImageBridge node is the fallback for the same endpoint.