GPT Image 2.5 character drift: what OpenAI says and what to send
OpenAI's image guide says GPT Image may struggle to keep a recurring character the same across generations. What to send on Sume to reduce drift.
Yes, GPT Image can drift. OpenAI's image generation guide lists consistency as a limitation: the model can produce consistent imagery but "may occasionally struggle" to keep recurring characters or brand elements the same across multiple generations. The practical fix is to stop describing the character from scratch each time and send the same reference image with every request, which on Sume means input_references on POST /v1/images.
OpenAI's statement is from its Image generation guide, read 2026-10-02. Sume's request fields are from the Sume Image API page, read the same day.
What exactly does OpenAI say?
The guide's Limitations list has four items: latency (complex prompts may take up to 2 minutes), text rendering, consistency, and composition control. The consistency item says that while the model can produce consistent imagery, it may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations.
That is a vendor caveat, not a measurement. It does not say how often drift happens, so treat any single series as something to check by eye.
Which Sume request reduces drift?
Use the same reference image in every call and put the unchanging identity words in one reusable block. Sume's catalog advertises up to 16 input_references for openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst, so a face shot, an outfit shot and a product shot can all ride along.
Edits behave differently from fresh generations: the docs say to prefer aspect_ratio: "auto" on edit calls so the output follows the reference, and that leaving it out is not the same thing.
| Hold fixed | How on Sume | Why |
|---|---|---|
| Face and outfit | Same image in input_references | The model sees the same pixels each time |
| Identity wording | One prompt block, pasted unchanged | Only the scene changes between calls |
| Frame shape | aspect_ratio: "auto" on edits | Output follows the reference shape |
| Model id | One pinned id such as openai/gpt-image-2.5 | Avoids a different model mid-series |
Does Sume keep state between calls?
No. Each POST /v1/images call carries its own prompt and references, and the Image API page documents no conversation or response-id field. OpenAI offers multi-turn editing in its Responses API, which Sume's Image API does not mirror. Re-send the reference on every call.
Also note that seed is in the Sume schema but no model advertises it, so a request that sends one returns 400 unsupported_parameter. You cannot pin a seed to get repeatable characters.
How do I check a series for drift?
Generate the frames, put them side by side, and compare the parts that must not move: face shape, hair, logo placement. If frame four drifts, re-run it with the first frame as the reference rather than the third, so errors do not stack up from one edit to the next.
A slow high-quality frame may return 202 with a job envelope; read the images from GET /v1/jobs/{id}/result as the Jobs and results page describes.
Sources
Related posts
More in Models
- gpt-image-2.5 smallest custom size: the 655,360-pixel floor
gpt-image-2.5 on Sume rejects images under 655,360 pixels. Why 512x512 and 768x768 fail, which small sizes pass, and the other edge rules to check first.
- GPT Image 2.5 misspells a brand name: spell it letter by letter
OpenAI's prompting docs say to quote the text and spell tricky words letter by letter. The prompt to write, and how to check the result on Sume's image API.
- GPT Image 2.5 text in the wrong place: fix it with a mask edit
OpenAI says GPT Image 2.5 can still misplace text. Mask only the text zone and re-render that region through Sume's mask_url edit instead of the whole image.
- grok-4.5 as a Sume model id runs GPT-6 Sol, not Grok: the alias list
Sume treats grok-4.5, composer-2.5, composer-2.5-fast and sume-1.0-fast as old Cursor-era ids that run GPT-6 Sol. How Grok 4.7 is gated and how grok-4.6 moves.
Written by Sume