GPT Image 2.5 Sunburst · Edit

Sunburst is the quality-first half of the GPT Image 2.5 pair, and the edit node is where that matters most. Small type on a label, a logo on a curved surface, a product shot that has to survive being blown up - those are the jobs where the cheap tier's softness gets caught, and where 4K resolves it. Add a real mask socket and sixteen reference slots and you've got a retoucher that costs cents per attempt.

What it is

PopNow_openai_gpt_image_2_5_sunburst_edit calls OpenAI's GPT Image 2.5 Sunburst edit endpoint through the PopNow pack. Starting at about 0.27 credits ($0.0027) per image - the low tier - scaling up with resolution, quality and output count. Cloud-only, no weights, prompts and source images leave your machine, filter included. That's the standing trade for a model you can't download.

It's the same input set as the Flare edit node, and the same width: up to sixteen sources. The difference is what the model is optimised for. Flare is fast, Sunburst is detailed.

How it works

Attach sources to image_urls (an Autogrow socket - each connection adds a slot, up to sixteen), describe the change, and the node uploads your images (PNG, JPEG fallback above 10 MiB), quotes the run, and downloads the result as an IMAGE batch. Uploads are cached by content hash for 12 hours, so iterating on wording doesn't re-send the same product shot six times.

The alpha output comes back alongside images on every run - the result's transparency as a MASK. That's not decoration. Paired with background set to transparent, it turns the node into a cut-out generator: subject on a real alpha channel, ready to composite.

The inputs

• prompt - say what to change and what must stay the same. This model follows instructions literally, so explicit preservation clauses ("keep the label text and the bottle shape, change only the backdrop") do real work.
• image_urls - one to sixteen sources, in the order your prompt refers to them.
• mask_url (optional) - a PNG where transparent pixels mark the region to repaint in the first source image. Feed it an actual alpha channel, not a black-and-white matte.
• input_fidelity - low by default; high preserves more of the source. For packaging, logos and faces, use high. For "restyle this whole thing", low is fine and often better.
• aspect_ratio - 1:1 default. The output does not follow your source, so set this on purpose.
• resolution - 1k, 2k, 4k.
• quality - high by default, with tiers above it. This is the axis that buys you the small details Sunburst is known for.
• num_outputs - 1 to 10, billed per image.
• background - auto default, plus transparent and opaque.
• output_format, output_quality - png default; JPEG/WebP quality 0–100, or -1 for the model's choice.

Outputs: images, alpha, task id.

The 4K workflow

Do the composition work at 1K. Iterate the prompt ten times, get the framing and the preservation right, then raise resolution and quality once for the take you're keeping. Every tier multiplies the per-image rate, and num_outputs multiplies again - a four-up 4K batch is not "a few cents", it's a real line item.

Then check the result at 100% on the thing you actually care about: the small text, the seam at the mask edge, the reflection. Sunburst's advantage over the cheap tier lives exactly there, and if you don't look, you've paid for detail you didn't check.

Install

Manager → search PopNow → Install, or:

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

Restart, add a key from popnow.ai/api/console/keys in Settings > PopNow, restart again. The pack installs no Python dependencies of its own.

Traps

Mismatching the mask is the classic one: mask_url applies to the first image only, so if you've got three references and the mask was built for the third, you'll edit the wrong picture. PopNow content_rejected is the provider's filter and is refunded. PopNow 40201 is credits. And a mask that's fully transparent means "repaint everything", which will look like the edit ignored your instruction - check the mask before blaming the model.