GPT Image 2.5 Sunburst · Text to Image

If the deliverable has to say something - a poster headline, a menu, a packaging label, a fake app screen - resolution is what separates a legible render from a mushy one. Sunburst Text to Image goes to 4K and is the quality-first end of OpenAI's 2.5 generation. It's also the one I'd finish on rather than explore with, and that distinction is most of the workflow advice here.

What it is

A hosted text-to-image node for GPT Image 2.5 Sunburst in the PopNow pack. Seven inputs, no source images, no mask, cloud inference only. Starting price is about 0.27 credits ($0.0027) per image at the low tier, and the price climbs with resolution, quality and how many outputs you ask for.

The family's pitch is in-image text that holds up. That's less of a novelty than it used to be, but it's still the reason people put up with a closed model: you are never going to get a diffusion checkpoint to render a nine-word headline with correct spelling on the first try and keep it through a re-roll.

How it works

Prompt in, settings in, IMAGE batch out. Behind that: PopNow quotes the run before anything is held, renders on OpenAI's side, and the node downloads and decodes the returned files. With num_outputs above one, all the results land in a single batch - Save Image writes one file each, so you don't lose the extras.

Worth internalising: there's no seed and no input to cache. Every queue is a fresh, fully-priced render, which is why the drafting discipline below matters more here than on the video nodes.

The inputs

• prompt - describe the image. For text renders, put the exact words in quotes and say where they go and how big: a headline set small in the corner is where these models fail, and telling them it's small doesn't fix it. Give text room.
• aspect_ratio - 1:1 default, with the full spread up to 16:9, 9:16 and the extreme banner ratios.
• resolution - 1k, 2k, 4k. The reason you're here.
• quality - high by default; higher tiers add detail, time and cost.
• num_outputs - 1 to 10, each billed separately.
• output_format - png default, plus jpeg and webp. Keep PNG when text will be scaled.
• output_quality - 0 to 100 for lossy formats, -1 to leave it to the model.

Outputs: images and task id.

Draft on the cheap tier, finish here

This is the single habit that makes the GPT Image nodes affordable. Explore composition and wording on the Flare text-to-image node - same pack, same catalog, lower cost per attempt - until the prompt is settled. Then move the final prompt to Sunburst, raise resolution and quality, and generate one or two images.

The reason it works is that the two nodes take near-identical inputs, so moving a prompt between them is a copy-paste. The reason it's necessary is that the quality tiers are multiplicative: high quality at 4K across six outputs is not a cheap experiment, whatever the badge's "from" figure says.

Install

Manager → search PopNow → Install, or:

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

Restart, then add a key from popnow.ai/api/console/keys via Settings > PopNow and restart once more so the node list rebuilds from your account. Zero dependencies beyond ComfyUI itself.

Troubleshooting

PopNow 40201 is an empty wallet; PopNow content_rejected is the filter and is refunded on failure. PopNow 40004 means that combination of settings isn't servable and names the field to change - usually a very narrow aspect ratio at a high resolution tier. If the node isn't in your list at all, restart ComfyUI: the catalog is fetched at startup and cached, and a stale node list is nearly always a missing restart rather than a missing model.

One expectation to set: this is not a local model and never will be, so there's no abliteration, no LoRA, no negative prompt to fight a refusal with. Useful to know before you spend an evening trying.