GPT Image 2.5 Flare - Text to Image

What "Flare" means here

Flare is KIE's catalogue name for one of the GPT Image 2.5 variants it proxies. Don't go looking for an OpenAI model card under that name - this is a reseller SKU, and the pack turns each SKU into its own ComfyUI node. Yours lives in KIE Next/Image/OpenAI/GPT Image and does text-to-image.

What you get over the older GPT Image nodes is resolution: this one offers a real 1K / 2K / 4K picker and a long aspect-ratio list. The 2K/4K options are the point. A poster, a product shot, packaging art, anything with small typography you intend to print - that's the use case. Local models rival this quality fine; what they don't do is hand you a clean 4K frame from a paragraph of direction without a captioning workflow bolted on.

Nothing runs locally. No weights, no VRAM, no free tier - each run is a metered call through KIE.

How the pack drives it

Node classes are generated at startup from KIE's API documentation, so the class id is derived: KIE_Next_gpt_image_2_5_flare_text_to_image_8d103001, where the trailing hash comes from the model id and its documentation URL. That's deliberate - the ID stays stable across pack releases instead of changing every version.

The run path is the pack's standard market flow: POST to KIE's createTask with the model id gpt-image-2-5-flare-text-to-image and your inputs, poll recordInfo until the task reports success, download the returned image, hand you a tensor. Transient errors get retried with exponential backoff, and the pack waits out the short window where a task is marked successful before its URL is readable - worth knowing, because it means a slow provider doesn't cost you a paid generation.

Inputs

prompt is required, multiline, and tolerates long prose - the documented limit is 20,000 characters. The default, "a cinematic night city poster with neon reflections on a rainy street," is fine but generic. For this node specifically, put the text you want rendered in quotes inside the prompt and keep it short; long strings of lettering are where every model in this class starts improvising glyphs.

aspect_ratio defaults to 3:2, with auto, 1:1, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9 and more available - thirteen options in all. The tooltip carries a caveat the list doesn't: the unusual ratios 27:16, 16:27, 9:8 and 8:9 are 1K only, and 2K/4K work on the other ratios. If you pick an exotic frame and 4K, expect trouble; match a mainstream ratio to a high resolution.

resolution defaults to 1K here, with 2K and 4K above it. This is the setting you're actually here for. Go up a step and the call costs more and takes longer; start at 1K while you're finding your composition, then re-run the winner at the resolution you need.

background defaults to transparent with opaque and auto alongside. That default is worth thinking about rather than accepting. If you want a finished picture, transparent is not what you want - pick opaque. If you genuinely want a cutout to composite, transparent is right, but the rest of your graph then has to carry alpha properly: PNG on the save side, and be aware that anything downstream expecting three channels will flatten it against something.

timeout_seconds defaults to 1200 (30–7200, step 30). 4K generation is slow; don't shrink this unless you want early failures.

callback_url is for KIE to push results to your own server. Leave it empty - the node polls.

There's no seed and no negative prompt, so there's no reproducibility and no way to subtract. You iterate by editing what you got.

Outputs

image goes to Save Image, a preview, or an upscaler. url is the KIE-hosted file, all_urls_json is every URL the task returned (the image socket only carries a single tensor from the first usable URL, so if a job ever returns multiple images the rest live only in this JSON), task_id is the identifier to quote in a support ticket, raw_json is the provider's complete response for debugging, and credits_consumed / credits_left report the cost - the same two numbers also render on the node as a receipt after each successful run, which is how you catch a loop quietly burning credit.

Install

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

Restart ComfyUI, then set the key once at Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. The pack validates it against KIE before overwriting anything and only ever shows a masked fingerprint afterwards; KIE_API_KEY in the environment wins over the saved key for headless installs. Dependencies are requests, Pillow, numpy, PyYAML - no CUDA, no model downloads. Registry/Manager publication is described as "prepared," so if Manager's search doesn't surface the pack, clone it.

Notes and pitfalls

Ration your 4K calls. There's no local fallback and no free tier, so twenty variations at 4K is twenty expensive images. Compose at 1K, commit at 4K.

Straight from the docs. Every default in this node - the prompt text, the ratios - is lifted from KIE's published example. It tells you what the reseller tested, not what's best.

The filter is the provider's. Refusals come from the model behind the endpoint and can't be configured away on the node.

Node missing from the menu? The pack builds most nodes from a live catalog sync on first launch (and re-checks the catalogue periodically). Restart once, or run the Refresh Model Catalog node.