GPT Image 2.5 Flare · Text to Image

OpenAI's image models have one unambiguous selling point over the open field - they write text into a picture and it comes out spelled correctly - and one long-running complaint: moderation that turns on a dime. Flare is the speed-first model of the 2.5 pair, from about $0.001816 an image, with five quality tiers, 1K to 4K output and fifteen aspect ratios from 1:3 to 3:1.

Where this fits

Picture the workflow: a poster, a menu, a shop sign, an infographic with real words in it. Flux and the SDXL family can often get the vibe and rarely get the lettering, which is exactly why people keep a closed model within reach. The GPT Image line has been the reference for that since GPT Image 1 in April 2025, and by OpenAI's own framing the primary surface is ChatGPT - the enthusiast community mostly doesn't use it for serious work because the censorship swings unpredictably. That's the trade: capability you can't get locally, at a per-call price, with a filter you don't control.

This node makes that call from inside ComfyUI. It's a wrapper: the model runs on SpicyAPI's servers, so no GPU and no downloads, and the prompt plus any inputs leave your machine on every run.

Inputs and outputs

• prompt - describe the image. Be explicit about text you want rendered; that's what you're paying for.
• aspect_ratio - fifteen ratios, from 1:3 through to 3:1, defaulting to 1:1. That range is unusually wide and useful for banners and vertical formats.
• resolution - 1k, 2k or 4k.
• quality - five tiers from low to max. Higher tiers add detail, time and cost; for a text-heavy graphic, quality is where your legibility budget goes.
• num_outputs - 1 to 10, each billed separately. This is the field that turns a cheap-looking rate into a real bill.
• output_format - png by default, JPEG and WebP also supported.
• output_quality - −1 hands compression to the model; 0–100 for the lossy formats only.

Outputs are images and task id. No alpha out on this variant (the edit nodes have one); wire images into Save Image or a preview.

Picking Flare over Sunburst

They're a pair and they share almost every input. Flare is the speed-first one; Sunburst is the quality-first one. For prompt exploration - where you're going to throw away three of four results - Flare is the sensible start, and for the final take that will sit on a poster, try Sunburst on the same prompt and compare. Both are listed at the same headline rate, so the decision is about time and quality, not price.

Worth knowing before you plan a pipeline around it: the KB's closed-model catalogue tracks GPT Image 1 and 1.5 and hasn't been updated for the 2.x generation, so treat the model-specific behaviour claims here as the node description's own, not community-verified. Test on your own prompts rather than trusting anyone's summary.

Installing it

ComfyUI Manager → search SpicyAPI → install → restart. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, and restart once so the node list is rebuilt from your account catalogue. If the model you want isn't in the menu, that restart is usually the whole fix - the list is generated at startup. Headless machine: SPICY_API_KEY, and the pack honours HTTPS_PROXY if you're behind one. There's nothing else to install; the pack declares no dependencies at all.

The bill, briefly

Set Max cost per run in Settings > SpicyAPI before your first run. The node quotes every job, compares the worst case against your ceiling, and raises before creating a task if it's over - so the seatbelt exists and costs you nothing to wear. Note the badge on the node is a rate, not a total: the pack deliberately doesn't fake an estimate, since resolution tier and output count change the real number. The charge gets printed on the node when the run finishes, and a content_rejected result means the model said no - typically with no charge, per the model page.