GPT Image 2.5 on Visiomake: We Measured What It Changes When You Edit a Render

GPT Image 2.5 — the API side of OpenAI's ChatGPT Images 2.5 release — is now live on Visiomake. It replaced GPT Image 2 on 16 September 2026 in the AI Image Generator and in Moodboard to Render, at the same 30 credits per image.

OpenAI's pitch for this release is editing: change what you asked for, leave the rest of the image alone. For AI render editing — revising an image a client has already seen — that is the only claim that matters, so instead of repeating it we tested it. We took three interior renders, gave each one a single revision of the kind a client actually sends, and ran the same instruction through the old GPT Image 2, the new GPT Image 2.5 and Nano Banana Pro. Then we measured how much of the image changed outside the thing we asked to change.

The short version: GPT Image 2.5 drifted less than GPT Image 2 on two renders and matched it on the third — and it finished each edit in about 41 seconds instead of about 142. The speed is the bigger upgrade. None of the three models left the rest of the frame truly untouched, and the places where they drift are predictable. Details, numbers and limits below.

What changed on Visiomake

OpenAI released ChatGPT Images 2.5 on 8 September 2026. In the API it is not one model but two: GPT Image 2.5 Flare, tuned for latency, and GPT Image 2.5 Sunburst, which OpenAI describes as built for workflows that need tighter control across edits. Visiomake runs Sunburst. Most image work on the platform starts from a reference — a render, a moodboard, a product photo — so keeping that reference intact is worth more to our users than shaving seconds off a text-only generation.

• Where it is: the AI Image Generator (text-to-image, with up to 16 reference images) and Moodboard to Render (1–5 references), listed as GPT Image 2.5.
• Price: 30 credits (€0.30) per image — unchanged from GPT Image 2.
• Output: 1:1, 16:9 and 9:16 formats; square output is 1024×1024 px and landscape is 1536×1024 px. Reference images can be PNG, JPEG or WebP.
• GPT Image 2 is no longer in the model picker. Everything you generated with it stays in your History.

There is no separate “edit” mode to look for. To revise a render with GPT Image 2.5, open the AI Image Generator, attach the render as a reference image, and write the change as an instruction. That is exactly what the test below does.

The test: one client revision, three models

We generated three square interior renders with Nano Banana Pro at 2K, then wrote one revision for each — an object swap, a material change, and a replacement on a wall:

• Living room: replace the plain cream wool rug with a terracotta-and-navy Persian-style rug of the same size.
• Kitchen: change every cabinet front, wall units and island, from matte white to dark walnut veneer.
• Bedroom: replace the round black-framed mirror above the dresser with a tall arched mirror in a brushed-brass frame.

Every instruction ended the way a careful brief does, by listing what must not move. The kitchen one read, in full: “Change all the kitchen cabinet fronts, on the wall units and on the island, from matte white to dark walnut veneer with a satin finish and vertical grain. Keep everything else exactly as it is: the camera angle, the worktop, handles, oven, pendant lights, flooring, walls, window, the objects on the island and the lighting.”

Each instruction went, word for word, to three models: GPT Image 2 and GPT Image 2.5 Sunburst (API id gpt-image-2.5-sunburst) through OpenAI's API with the same parameters Visiomake sends (1024×1024, quality high), and Nano Banana Pro through Visiomake itself at 2K. That is nine edits, one run each.

To score them we drew a generous outline around the region each instruction targeted, scaled every result to 1024×1024, and compared it with the original render outside that outline. Three figures per edit: the average per-pixel difference on a 0–255 scale, the share of pixels whose brightness moved by more than about 10% of the full brightness range, and structural similarity (SSIM, where 1.0 means identical). Lower is better for the first two, higher for the third.

* Nano Banana Pro was timed through Visiomake's job queue, from submission to finished asset. The two GPT models were timed as direct API calls. The GPT figures compare cleanly with each other; the Nano Banana Pro times are indicative only.

What the numbers say

1. The real upgrade is speed

GPT Image 2 took 138–149 seconds per edit; GPT Image 2.5 took 37–45 — about 3.5 times faster on identical requests. That is a bigger gap than OpenAI itself advertises (it claims up to 50% lower latency, and for the Flare model, not Sunburst), and it comes from three calls on one day, so treat the exact multiple loosely. The direction matches what we saw in production: the median GPT Image 2 generation on Visiomake ran 157 seconds, long enough to break the rhythm of a working session. A revision loop where each attempt costs 40 seconds instead of two and a half minutes is a different way of working — you can afford a second and a third attempt while the client is still on the call.

2. It drifts less than GPT Image 2 — consistently, not dramatically

GPT Image 2.5 came out ahead of its predecessor outside the edit region on all three renders and all three measures — but on the rug the gaps are so small (1.64% against 1.65%) that it is a tie, not a win. Averaged, the share of pixels that changed where nothing should have fell from 2.67% to 1.82%. The clearest gap was the bedroom: GPT Image 2 nudged the dresser's drawer lines and the folds of the bedding while it swapped the mirror (SSIM 0.834), and GPT Image 2.5 held them noticeably better (0.890). In the kitchen GPT Image 2 also did something the other two did not: it re-skinned the stainless filler panel under the microwave in walnut, part of an appliance we had asked it to keep. This is a real improvement, and it is not night and day — at normal viewing size both results would pass as “the same room.”

3. GPT Image 2.5 vs Nano Banana Pro: close, with one exception

On the rug and the mirror, GPT Image 2.5 and Nano Banana Pro landed within about half a point of each other on both pixel measures. On the kitchen they did not: Nano Banana Pro changed 6.31% of the pixels outside the cabinets against 0.16% for GPT Image 2.5. Nothing broke — it darkened the rest of the room, the floor most of all, by about 10 brightness levels out of 255 on average, where GPT Image 2.5 did not shift it at all. You can argue Nano Banana Pro is the one being physically honest: walnut bounces far less light into a room than white lacquer, and a real re-render would darken too. But if the approved image's exposure is part of what the client approved, it is a change you would have to grade back, and GPT Image 2.5 is the model that did not make it. Note also that the base renders were Nano Banana Pro's own output, which if anything should have favoured it.

No model left the rest of the render untouched

This is the finding worth remembering whichever model you pick. An instruction edit is not a Photoshop layer: all three models re-render the entire frame and try to make it match the original. They are very good at it — every book on the living-room shelf, every lemon on the kitchen island came back in place — but “very good” is not “identical.” In our heatmaps, apart from Nano Banana Pro's exposure shift in the kitchen, the drift collected in the same kind of place every time: fine, repeating textures. The jute rug and the rumpled linen in the bedroom, the olive tree's foliage in the living room. Hard-edged furniture, walls and windows barely moved on GPT Image 2.5 and Nano Banana Pro; GPT Image 2 was the one that also shifted the dresser's drawer lines and, in the kitchen, re-skinned an appliance panel.

Two practical consequences:

• Check textures, not furniture, before you send a revision. If the client signed off a specific fabric or a stone with distinctive veining, zoom into it on the edited version. That is where a silent change will be.
• If the rest of the frame must stay as it is, mask the change instead. A masked inpaint in the Render Editor only asks the model to fill the area you paint, which is the safer route for a small, local fix on an approved image. We covered that trade-off in AI inpainting versus re-rendering for visualization revisions. Instruction edits earn their place when the change is spread across the image — every cabinet front in a kitchen is miserable to mask and trivial to describe.

Which model to pick for a revision

Both are in the same model picker, so the choice is per image, not per subscription.

• Pick GPT Image 2.5 when the brief is “change this, keep that” and the approved image's lighting and exposure are part of what was approved. It had the lowest or tied-lowest pixel drift on every render, the one clear gap — the kitchen — went its way, and at 30 credits it is the cheaper of the two.
• Pick Nano Banana Pro when you need the pixels. GPT Image 2.5 tops out at 1536×1024 on Visiomake; Nano Banana Pro generates at 2K for 35 credits and, on a paid account, 4K for 63, which matters for print and for crops — we went through the arithmetic in when Nano Banana Pro's 4K output is worth paying for. It was also the fastest model here.
• Either way, write the instruction like a brief. Name the one thing that changes, describe the replacement in material terms (species, finish, grain direction, frame metal), then list what stays: camera, furniture, lighting, the objects on surfaces. If revisions are a regular part of your week, how to handle client change requests on a final render has the longer workflow.

A GPT Image 2.5 result you are happy with is still a 1024- or 1536-pixel image. For anything beyond a screen-sized preview, finish with an upscale — the same logic as the low-resolution-plus-upscale render workflow.

What we did not test

This was a small test, and it should be read as one.

• One run per model per render. These models are not deterministic; run the same edit again and the figures will move. We measured no variance, so treat gaps of a few tenths as ties. The direction of the GPT Image 2 to 2.5 result was the same on all three renders, which is why we are comfortable reporting it.
• Three interiors, square format only. No exteriors, no 16:9, no 9:16.
• Single-reference edits only. Moodboard-style generation from several references is a different task and we did not score it.
• Quality high only — the setting Visiomake sends. GPT Image 2.5 Flare was not tested at all.
• One small deviation from the product path: our script sent OpenAI a PNG re-encode of each base render, where Visiomake would send the stored JPEG.
• The edit outlines were drawn by hand with a margin. In the kitchen, some darkening outside the cabinets is physically correct, because darker fronts reflect less light into the room.
• Pixel drift is not the same as quality. All nine edits did what was asked, and all nine would survive a client review at normal size. We measured how faithfully the untouched areas were preserved, not which result looks best.