GPT Image 2.5 Flare · Text to Image

Seven inputs, no source images, no mask. This is the plainest node in the GPT Image corner of the PopNow pack - you type what you want, you get a picture, and it starts around 0.27 credits ($0.0027) per image. Flare is the speed-first 2.5 variant, so treat it as the drafting tool rather than the final-render tool.

What it is

A hosted text-to-image endpoint for GPT Image 2.5 Flare, exposed as a ComfyUI node by the PopNow pack. GPT Image is OpenAI's image line and its reputation rests on two things: instruction-following that actually reads your prompt as instructions, and in-image text that's legible. If you need a poster with a headline on it, or a UI mock with readable labels, this is the family that made that stop being a joke.

There is no local path to this model - the weights aren't yours and never will be. The node is the door: per-call pricing, prompts leave your machine, the provider's filter applies. That's the deal.

How it works

The prompt and settings go out, PopNow renders, the node downloads the returned files and decodes them into an IMAGE batch. num_outputs above 1 returns several images in that one batch, which is why Save Image writes a file per item rather than overwriting.

Nothing is cached here in any useful sense - text-to-image has no input to reuse, so every queue is a purchase. The node quotes first and holds at most the quoted amount, and the price badge updates as you change resolution, quality and output count, which is worth watching before you hit queue on a 10-image batch.

The inputs

• prompt - describe the image you want. Long, specific prompts work; this model rewards being told what's in frame, what the light is doing, and what the text should say.
• aspect_ratio - 1:1 by default, with a wide spread including 16:9, 9:16, 4:5, 5:4 and much narrower ratios like 1:3 and 3:1. The extreme ones are for banners and mobile crops, and they're priced differently.
• resolution - 1k, 2k, 4k.
• quality - high by default, and the tiers above it add detail, time and cost. This is the first knob to turn down when you're exploring.
• num_outputs - 1 to 10. One request, one bill per image returned.
• output_format - png by default; jpeg and webp available. Keep PNG if there's text you'll want to scale.
• output_quality - 0 to 100 for JPEG and WebP; -1 leaves it to the model. Irrelevant for PNG.

Outputs: images and task id.

What it's good at, and what it isn't

Good: typography, signage, packaging, flat illustration, anything where the brief is "convey this specific information". It reads a lengthy paragraph and follows the constraints in it, which is unlike most diffusion models that quietly ignore the third sentence.

Not good: fine art grain, painterly texture, anything where you want a specific aesthetic lineage rather than a clean commercial image - and anything the filter dislikes, which you can't argue with because there are no weights to patch. The community's allergy to this model is mostly about that last part plus the per-call cost, and neither complaint is wrong.

Draft cheap, finish sharp

Draft at 1K with quality at its default and num_outputs at 1 until the prompt is right. Then raise resolution and quality for the one you keep. The per-image rate is genuinely low, but it multiplies across every axis you turn up, and a habit of queuing four variations at 4K adds up faster than you'd think.

If you need an edit - change this, keep that - use the edit node instead of trying to talk this one into a fix-up. Text-to-image re-rolls everything, including the parts you liked.

Install

Manager → search PopNow → Install, or:

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

Restart, paste a key from popnow.ai/api/console/keys into Settings > PopNow, restart again. No Python packages to add; the pack's dependency list is empty and it runs on ComfyUI's bundled node API.

Errors worth knowing

PopNow 40201 is an empty wallet. PopNow content_rejected is the provider's filter, refunded on failure. PopNow 400 naming a field means the model won't serve those settings together - the model page linked in the node description lists what it accepts. And if you're on a shared box, remember the key lives in user/popnow/config.json or in POPNOW_API_KEY, never in the workflow - so a shared graph won't leak it, but a shared machine might bill the wrong person.