GPT Image 2.5 Sunburst - Text to Image

Meet the sibling

Sunburst and Flare are two variant names in KIE's GPT Image 2.5 catalogue - reseller SKUs, not names you'll find on an OpenAI model card. The pack turns each into its own node, and this one is the text-to-image half of Sunburst, sitting in KIE Next/Image/OpenAI/GPT Image.

The differences from its Flare sibling are all in the defaults: Sunburst starts at 4:3 and 2K with an opaque background, where Flare starts at 3:2, 1K, and transparent. Same controls, different starting point, and the defaults say Sunburst is aimed at finished pictures rather than cutouts. That's usually the right starting point for photorealism work and wrong for compositing, so pick your node by what you intend to do with the pixels afterwards.

As with everything in this pack: no local model, no weights, no free tier. Each run is a metered call to KIE that forwards the job to the provider.

How it runs

The pack builds node classes at startup from KIE's API documentation, which is why the class name looks generated - KIE_Next_gpt_image_2_5_sunburst_text_to_image_1fd2c253. The hash comes from the model id and its docs URL, not the pack version, so the ID is stable across releases.

Execution is the standard market flow: POST to KIE's createTask with model id gpt-image-2-5-sunburst-text-to-image and your inputs, then poll recordInfo until the task reports success, then download the image and return a tensor. Transient failures - rate limits, gateway errors, dropped sockets - are retried with backoff, and a short grace period handles the case where a task is marked successful a moment before its result URL is actually fetchable.

Inputs

prompt is required, multiline, and accepts long instructions, up to 20,000 characters. The default - "a cinematic night city poster with neon reflections on a rainy street" - is KIE's documentation example, and it's a decent illustration of the node's strength: a scene described like a sentence, not a tag pile. If you want literal text in the image, wrap it in quotes and keep it brief.

aspect_ratio defaults to 4:3 and offers thirteen values including auto: 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 27:16, 16:27, 9:8, one more, and auto. The tooltip flags that 27:16, 16:27, 9:8 and 8:9 are 1K only, so exotic frames and 2K/4K don't mix - pick a mainstream ratio if you're going up in resolution.

resolution defaults to 2K, with 1K and 4K around it. That middle default is a sensible compromise and it's why this node exists as a separate entry rather than just a preset on Flare: at 2K you get noticeably cleaner detail than 1K without the 4K cost. Step down to 1K to iterate, up to 4K for the keeper.

background defaults to opaque with transparent and auto available. Opaque is the sane default for a finished image. Transparent is the one to reach for deliberately - for a subject you intend to cut out and place yourself - and it means your downstream nodes have alpha to deal with, so save PNG and check anything that expects three channels.

timeout_seconds defaults to 1200 seconds, adjustable from 30 to 7200 in 30-second steps. 2K and 4K jobs are slow; the default exists for a reason.

callback_url lets KIE push the result to a server you run instead of being polled. Leave it blank in ComfyUI.

No seed, no negative prompt - so no reproducibility and no subtraction. Iterate by editing output.

Outputs

image is the tensor: Save Image, preview, upscaler, or an edit node for round two. url is the KIE-hosted file, all_urls_json contains every URL the task returned (only the first usable one becomes the image tensor, so anything extra is only visible here), task_id is what you quote to support, raw_json is the provider's full response, and credits_consumed / credits_left are the spend - mirrored onto the node as a receipt after a successful run, which is the fastest way to notice a workflow that's costing more than you assumed.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r requirements.txt

Restart ComfyUI, then paste your key once at Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. It's validated before it's saved, and afterwards only a masked fingerprint appears; KIE_API_KEY in the environment takes precedence for headless installs. Requirements are requests, Pillow, numpy, PyYAML - nothing to compile and no model downloads. The README offers a Windows one-click installer and packaged ZIPs in dist/, and describes Registry/Manager publication as "prepared," so if Manager's search doesn't find the pack, clone it directly.

Things to watch

Compare against your local models honestly. A local Flux or SDXL run costs you electricity and nothing else; this costs per image. Use it where the hosted model wins - long natural-language direction, clean typography, fast iteration with zero VRAM - and keep the bulk of a batch local.

Aspect ratio and resolution interact. The 1K-only ratios in the tooltip are a hard constraint, not a hint.

The catalogue moves. KIE adds, renames and retires models, and the pack rebuilds its node list from their docs. A node vanishing after a refresh means the reseller changed something, not that you broke your install.