GPT Image 2.5 Flare: AI Image Editor Online

Everything below comes from VdoBloom's GPT Image 2.5 service, the Image Edit endpoint and the credit table. The headline is the 16-image reference cap, shared with Sunburst and above GPT Image 2 and GPT Image 1.5 (10 each).

The dollar columns use the Lite plan rate ($15 a month for 300 credits, $0.05 a credit) and the one-time Lite pack ($19.50 for 300 credits, $0.065 a credit). Pack credits never expire; plan credits reset each billing cycle.

VdoBloom's service code describes Flare as the fast, default variant and Sunburst as the variant that trades latency for edit precision and fidelity. For editing they are otherwise identical: 16 references, 13 ratios, 1K/2K/4K, 20,000-character instructions, and the same 3, 5 or 8 credits.

VdoBloom's own tools use both. The people-and-product tools (thumbnails, passport photos, re-angles, manga pages, ad creatives) run Flare; makeup, photo restore, relight and image translation offer Sunburst or Flare; Room Restyle and the die-cut sticker tool run Sunburst only. If an edit on Flare is close but not right, the same instruction on Sunburst costs the same, so trying both is a fair test.

All four edit from an instruction and carry the STRICT label. Flare's edit model has the highest reference cap of the four; Nano Banana 2.1 is the lowest price at 2K and 4K.

The price depends only on the resolution you pick, so the table shows a sensible resolution for each kind of edit. The reference count is what the job typically needs, up to the 16 the model accepts.

Background swap (1 image, auto, 2K): "Replace the background with a softly blurred city street at dusk, warm shop lights. Keep the person's face, hair, glasses, clothing and pose exactly as they are. Match the lighting on the person to the new background."

Product set (2 images, 1:1, 2K): "Image 1 is the product, a glass perfume bottle. Image 2 is the mood reference for colour and light. Place the bottle from image 1 on a wet black stone surface with the colour palette and light direction of image 2. Keep the bottle's shape, cap and label text unchanged."

Group composite (5 images, 16:9, 2K): "Images 1 to 4 are four friends; image 5 is a beach at sunset. Place all four people from images 1 to 4 standing together on the beach in image 5, left to right in upload order, natural sunset light on their faces. Keep each face and outfit as in their own photo."

Each one names the job of every image, states the change, and lists what must not change.

In the Image Edit tab you write the instruction yourself. In the tools, the prompt, ratio and resolution are set server-side; most run at 2K (5 credits an image).

Write the change first, then what must stay the same: "Replace the background with a softly lit grey studio sweep. Keep the face, hair, clothing, pose and camera angle unchanged." Saying what not to touch is the single most useful habit for edits.

With several references, give each image a job in the order you uploaded them: "Image 1 is the person, image 2 is the jacket, image 3 is the logo for the chest." With up to 16 inputs, numbering matters more than description.

Split multi-part edits into separate sentences rather than one long clause, so no part gets dropped. Quote any text to place in the image and ask for it to be reproduced exactly. Leave the ratio on auto to keep your photo's shape; choose a fixed ratio when the result must fit a placement. For transparent backgrounds, describe an isolated subject and render at 2K or 4K.

The four 1K-only ratios (27:16, 16:27, 9:8, 8:9) are lowered to 1K rather than rejected, and those edits are charged at the 1K price; pick a standard ratio when you need 2K or 4K. Each upload is capped at 10 MB in the Image Edit tab, and at least one image is required: for generation from text alone, use the Flare text-to-image model.

Instruction editors can change more than asked, especially faces in group shots; a keep-unchanged list and one change per run reduce that. Reports of soft or noisy results usually trace to an ambiguous instruction or a job that ran at 1K when 2K was needed.

The STRICT label means OpenAI's filters refuse restricted subjects; refused or failed edits are refunded. GPT Image 2.5 is proprietary and reached through VdoBloom's hosted route, with no downloadable weights.