Benchmark · run 6 October 2026

Ten edits in a row, four image models

Four image models, each editing its own picture ten times in a row with the same instructions and the same keep-clause. Every intermediate frame saved, every figure computed by script.

Results

10 edits in a row · one chain per model
Metric
GPT Image 2.5 FlareOpenAI
FLUX.3 ImageBlack Forest Labs
Seedream 5.0 FlashByteDance
MAI-Image-2.6 FlashMicrosoft
Cost per editMean over the ten edits, as reported by the endpoint
GPT Image 2.5 Flare $0.020 ten edits $0.20
FLUX.3 Image $0.024 ten edits $0.24
Seedream 5.0 Flash $0.018 ten edits $0.18
MAI-Image-2.6 Flash $0.023 ten edits $0.23
Median time per editWall clock, ten edits
GPT Image 2.5 Flare 47.0s ten edits 417 s
FLUX.3 Image 69.7s ten edits 663 s
Seedream 5.0 Flash 7.8s ten edits 79 s
MAI-Image-2.6 Flash 48.3s ten edits 408 s
Structure keptEdge overlap of the last frame with its seed, 0 to 1. Higher: more of the same building and camera
GPT Image 2.5 Flare 0.10
FLUX.3 Image 0.24
Seedream 5.0 Flash 0.10
MAI-Image-2.6 Flash 0.06
Drift from seedMean pixel difference, last frame against seed, 0 to 255. Lower: closer to the start
GPT Image 2.5 Flare 35.5
FLUX.3 Image 27.3
Seedream 5.0 Flash 48.4
MAI-Image-2.6 Flash 59.6
Change per editMedian pixel difference between consecutive frames. Lower: an edit touches less
GPT Image 2.5 Flare 7.6
FLUX.3 Image 3.3
Seedream 5.0 Flash 12.1
MAI-Image-2.6 Flash 34.3
Detail against seedSharpness of the last frame over the seed. 1.00 is equal; above it is more fine texture or noise, not necessarily detail
GPT Image 2.5 Flare 1.46
FLUX.3 Image 1.18
Seedream 5.0 Flash 1.80
MAI-Image-2.6 Flash 3.10
Sky shiftColour change of the top band, which no instruction names. 0 is none
GPT Image 2.5 Flare 21.4
FLUX.3 Image 4.3
Seedream 5.0 Flash 81.0
MAI-Image-2.6 Flash 40.6
Frame deliveredPixels
GPT Image 2.5 Flare 1536x864
FLUX.3 Image 1360x768
Seedream 5.0 Flash 1280x720
MAI-Image-2.6 Flash 1360x768 then 1365x768

One held. Three did not.

  1. FLUX.3 Image ended closest to its own seed on all three measures of departure: edge overlap 0.24 against 0.06 to 0.10 for the other three, drift 27.3 against 35.5 to 59.6, and a median edit that moved 3.3 grey levels against 7.6 to 34.3. The visual verdict, that FLUX.3 held up and the others degraded, is supported by these figures.
  2. The sky is the cleanest control, because no instruction touches it. It moved 4.3 for FLUX.3 Image and 21.4, 40.6 and 81.0 for GPT Image 2.5 Flare, MAI-Image-2.6 Flash and Seedream 5.0 Flash. In the frames this is the damage you can see: Seedream 5.0 Flash ends with a purple, high-contrast sky and GPT Image 2.5 Flare with a grainy one. For MAI-Image-2.6 Flash the figure is partly the building, whose roof climbs into the band as the camera closes in.
  3. The three failures are different. MAI-Image-2.6 Flash did not accumulate noise; it re-shot the picture from the first edit (edge overlap 0.14 after edit 1, 28.0 grey levels of change) and kept moving the camera. Seedream 5.0 Flash and GPT Image 2.5 Flare changed less per edit (12.1 and 7.6) but added texture that compounds: their detail ratio rose to 1.80 and 1.46.
  4. Sharpness alone would not have told the story. All four ended above 1.00 (1.18 to 3.10); FLUX.3 Image is nearest to its seed at 1.18. Above 1.00 here is grain, over-sharpened edges or a closer camera, not extra detail, so we do not read it as quality.
  5. One measure disagrees. The frame-margin colour shift, our general tone check, is 15.4 for FLUX.3 Image, larger than GPT Image 2.5 Flare (12.1). Most of it arrives at edit 5 (8.6 grey levels of change in one step, against 3.3 median), the edit that removes the tree standing in the left margin band. That is an instructed change landing in the margin, which is why we added the sky band as a cleaner control; we have not tried to correct for it.
  6. FLUX.3 Image is not the cheap or fast option here: 69.7 s median per edit and $0.024 per edit, against 7.8 s for Seedream 5.0 Flash (the fastest) and $0.018 for Seedream 5.0 Flash (the cheapest). It was the slowest of the four. Staying close to the seed was not free.
  7. Staying close is relative. FLUX.3 Image also moved: its final edge overlap is 0.24, not 1.00, and in the last frame the camera sits closer and the building reads simpler than in the seed. We did not measure camera movement on its own, but in the frames the camera moved in all four despite the lock in the keep-clause, by very different amounts.

The seed was generated once per model and reused for the chain; no chain was repeated. Frames differ in size between models. FLUX.3 latency for edits 1 to 8 comes from file timestamps, see the disclosure below.

Seed, edit 5, edit 10

Each model on its own seed
GPT Image 2.5 FlareOpenAI · 1536x864
GPT Image 2.5 Flare: the seed, a Balinese joglo villa with a rectangular pool
Seed
GPT Image 2.5 Flare after edit 5, no-frangi
Edit 5 · drift 25.9 · overlap 0.14
GPT Image 2.5 Flare after edit 10, channel
Edit 10 · drift 35.5 · overlap 0.10
FLUX.3 ImageBlack Forest Labs · 1360x768
FLUX.3 Image: the seed, a Balinese joglo villa with a rectangular pool
Seed
FLUX.3 Image after edit 5, no-frangi
Edit 5 · drift 24.6 · overlap 0.28
FLUX.3 Image after edit 10, channel
Edit 10 · drift 27.3 · overlap 0.24
Seedream 5.0 FlashByteDance · 1280x720
Seedream 5.0 Flash: the seed, a Balinese joglo villa with a rectangular pool
Seed
Seedream 5.0 Flash after edit 5, no-frangi
Edit 5 · drift 34.7 · overlap 0.16
Seedream 5.0 Flash after edit 10, channel
Edit 10 · drift 48.4 · overlap 0.10
MAI-Image-2.6 FlashMicrosoft · 1360x768
MAI-Image-2.6 Flash: the seed, a Balinese joglo villa with a rectangular pool
Seed
MAI-Image-2.6 Flash after edit 5, no-frangi
Edit 5 · drift 48.7 · overlap 0.06
MAI-Image-2.6 Flash after edit 10, channel
Edit 10 · drift 59.6 · overlap 0.06
GPT Image 2.5 Flare: all eleven frames
GPT Image 2.5 Flare, seed
Seed
GPT Image 2.5 Flare, edit 1, pool
Edit 1 · pool · 49.5 s
GPT Image 2.5 Flare, edit 2, tiles
Edit 2 · tiles · 17.1 s
GPT Image 2.5 Flare, edit 3, teak
Edit 3 · teak · 46.3 s
GPT Image 2.5 Flare, edit 4, bale
Edit 4 · bale · 47.4 s
GPT Image 2.5 Flare, edit 5, no-frangi
Edit 5 · no-frangi · 46.5 s
GPT Image 2.5 Flare, edit 6, pebbles
Edit 6 · pebbles · 47.6 s
GPT Image 2.5 Flare, edit 7, glass
Edit 7 · glass · 17.0 s
GPT Image 2.5 Flare, edit 8, second
Edit 8 · second · 45.8 s
GPT Image 2.5 Flare, edit 9, mahkota
Edit 9 · mahkota · 48.1 s
GPT Image 2.5 Flare, edit 10, channel
Edit 10 · channel · 51.4 s
FLUX.3 Image: all eleven frames
FLUX.3 Image, seed
Seed
FLUX.3 Image, edit 1, pool
Edit 1 · pool · 107.9 s
FLUX.3 Image, edit 2, tiles
Edit 2 · tiles · 63.6 s
FLUX.3 Image, edit 3, teak
Edit 3 · teak · 80.6 s
FLUX.3 Image, edit 4, bale
Edit 4 · bale · 45.0 s
FLUX.3 Image, edit 5, no-frangi
Edit 5 · no-frangi · 69.1 s
FLUX.3 Image, edit 6, pebbles
Edit 6 · pebbles · 83.6 s
FLUX.3 Image, edit 7, glass
Edit 7 · glass · 37.0 s
FLUX.3 Image, edit 8, second
Edit 8 · second · 70.4 s
FLUX.3 Image, edit 9, mahkota
Edit 9 · mahkota · 33.9 s
FLUX.3 Image, edit 10, channel
Edit 10 · channel · 71.4 s
Seedream 5.0 Flash: all eleven frames
Seedream 5.0 Flash, seed
Seed
Seedream 5.0 Flash, edit 1, pool
Edit 1 · pool · 8.5 s
Seedream 5.0 Flash, edit 2, tiles
Edit 2 · tiles · 8.3 s
Seedream 5.0 Flash, edit 3, teak
Edit 3 · teak · 7.4 s
Seedream 5.0 Flash, edit 4, bale
Edit 4 · bale · 7.7 s
Seedream 5.0 Flash, edit 5, no-frangi
Edit 5 · no-frangi · 8.4 s
Seedream 5.0 Flash, edit 6, pebbles
Edit 6 · pebbles · 7.7 s
Seedream 5.0 Flash, edit 7, glass
Edit 7 · glass · 8.0 s
Seedream 5.0 Flash, edit 8, second
Edit 8 · second · 7.6 s
Seedream 5.0 Flash, edit 9, mahkota
Edit 9 · mahkota · 7.8 s
Seedream 5.0 Flash, edit 10, channel
Edit 10 · channel · 7.4 s
MAI-Image-2.6 Flash: all eleven frames
MAI-Image-2.6 Flash, seed
Seed
MAI-Image-2.6 Flash, edit 1, pool
Edit 1 · pool · 52.9 s
MAI-Image-2.6 Flash, edit 2, tiles
Edit 2 · tiles · 55.7 s
MAI-Image-2.6 Flash, edit 3, teak
Edit 3 · teak · 18.0 s
MAI-Image-2.6 Flash, edit 4, bale
Edit 4 · bale · 18.5 s
MAI-Image-2.6 Flash, edit 5, no-frangi
Edit 5 · no-frangi · 49.9 s
MAI-Image-2.6 Flash, edit 6, pebbles
Edit 6 · pebbles · 47.4 s
MAI-Image-2.6 Flash, edit 7, glass
Edit 7 · glass · 47.3 s
MAI-Image-2.6 Flash, edit 8, second
Edit 8 · second · 19.5 s
MAI-Image-2.6 Flash, edit 9, mahkota
Edit 9 · mahkota · 49.2 s
MAI-Image-2.6 Flash, edit 10, channel
Edit 10 · channel · 49.3 s

Per step

Drift, edge overlap and detail, from run.json
Step-by-step figures for all four models
Edit
Change
GPT Image 2.5 Flare
FLUX.3 Image
Seedream 5.0 Flash
MAI-Image-2.6 Flash
1
pool
11.1 · 0.53 · 0.92
6.5 · 0.80 · 1.03
7.6 · 0.64 · 1.29
28.0 · 0.14 · 1.38
2
tiles
17.0 · 0.26 · 1.18
10.6 · 0.52 · 1.30
17.3 · 0.44 · 1.73
35.5 · 0.11 · 2.08
3
teak
20.8 · 0.21 · 1.27
13.9 · 0.50 · 1.37
25.4 · 0.32 · 2.01
41.5 · 0.10 · 2.13
4
bale
23.4 · 0.17 · 1.22
16.6 · 0.47 · 1.34
31.8 · 0.24 · 2.00
45.2 · 0.09 · 1.93
5
no-frangi
25.9 · 0.14 · 1.20
24.6 · 0.28 · 1.04
34.7 · 0.16 · 1.67
48.7 · 0.06 · 1.89
6
pebbles
27.9 · 0.12 · 1.37
26.4 · 0.26 · 1.16
39.0 · 0.14 · 1.76
54.6 · 0.05 · 2.16
7
glass
30.4 · 0.12 · 1.48
26.9 · 0.25 · 1.16
40.5 · 0.13 · 1.89
57.8 · 0.05 · 2.20
8
second
32.3 · 0.11 · 1.51
26.8 · 0.25 · 1.15
44.1 · 0.11 · 1.71
60.9 · 0.05 · 2.37
9
mahkota
33.8 · 0.11 · 1.45
26.9 · 0.25 · 1.17
46.4 · 0.10 · 1.76
61.4 · 0.06 · 2.14
10
channel
35.5 · 0.10 · 1.46
27.3 · 0.24 · 1.18
48.4 · 0.10 · 1.80
59.6 · 0.06 · 3.10
Each cell: drift from seed · edge overlap with seed · detail against seed.

A reference, not a contestant

An earlier chain, a different aspect
Muse Image 1.0, 3:2Meta · run published 1 September 2026Same seed prompt and the same ten edits, but a 3:2 frame and its own seed. Only its seed and its tenth frame are kept on this site, so no figure on this page is computed for it and it is not in the results table. Shown so the other four can be read against a chain with a different aspect ratio.
Reference chain, seed: the same villa from the same prompt
Seed
Reference chain, after ten edits
Edit 10

How it was run

Question

An architect’s session is a chain: change the pool, then the roof, then the timber, then add a pavilion, each edit applied to what the last one produced. We gave four public image models the same chain. Each generated its own picture of the same villa, then received the same ten instructions in the same order, every instruction applied to that model’s previous frame.

The question is how much of the picture is still there at the end: the building, the camera, the sky, the detail. Not whether one edit lands.

Models and access
  • GPT Image 2.5 Flare — OpenAI, via OpenRouteropenrouter.ai/api/v1/images · openai/gpt-image-2.5-flare, quality medium, aspect_ratio 16:9
  • FLUX.3 Image — Black Forest Labs, via OpenRouteropenrouter.ai/api/v1/images · black-forest-labs/flux-3-image, resolution 1K, aspect_ratio 16:9
  • Seedream 5.0 Flash — ByteDance, via OpenRouteropenrouter.ai/api/v1/images · bytedance-seed/seedream-5-0-flash, resolution 1K, aspect_ratio 16:9
  • MAI-Image-2.6 Flash — Microsoft, via OpenRouteropenrouter.ai/api/v1/images · microsoft/mai-image-2.6-flash, aspect_ratio 16:9
Seeds
One villa prompt, sent once to each model; the picture each returned is that model’s seed. Seeds were generated in a separate step and reused for the chains. No model edited another model’s picture.
Edits
Ten instructions, the same for every model, each applied to the model’s previous frame, every one followed by the same keep-clause. Verbatim below.
Aspect and size
16:9 requested from all four. Delivered sizes differ: GPT Image 2.5 Flare 1536x864; FLUX.3 Image 1360x768; Seedream 5.0 Flash 1280x720; MAI-Image-2.6 Flash 1360x768 / 1365x768. For measuring, every frame is resampled to its own seed’s grid at 768 px on the long edge.
Light and camera
Both are a control, not a variable. The seed prompt asks for flat overcast midday daylight, every edit repeats the lock, and no edit touches light, sky, weather or time of day. Regions no prompt names: the sky; the coconut palm grove behind the villa on the right; the jungle ridge on the horizon; the overcast daylight itself.
Measures
Computed by scripts/bench10-edit-chains-metrics.py over the saved frames: change per edit, drift, detail against the seed, edge overlap, margin tone and sky shift. All against the model’s own seed. The only judgement on this page is the reading of the frames in the findings.
Disclosure
  • Each model generated its own seed once, from the same prompt, in a separate step; the seeds were reused for the chains.
  • One chain per model, no repeats. Differences between neighbouring figures can be run-to-run noise.
  • Delivered resolution differs per model; every frame is resampled to its own seed's grid before measuring.
  • FLUX.3 steps 1-8: the run was interrupted before the raw output was written, so latency for those steps comes from file modification times and cost is the flat per-image price that every other recorded FLUX.3 call carried.
  • Sharpness, structure overlap and drift are measured against the seed, and the edits legitimately add and remove objects; they are comparable across models, not an absolute quality score.
Reproduce
Measurements, scripts and prompts are published under CC BY 4.0. The frames are outputs of third-party models, shown for comparison and commentary.
Scope

A dated snapshot of four public image models, executed on 6 October 2026. Providers update models under the same name; figures are left exactly as the run produced them. Providers update models without renaming them, so the date, endpoint and version of everything tested are stated on the page.

The ten edits, verbatim
1
Replace the rectangular swimming pool with an L-shaped pool that wraps around the near corner of the pavilion.form · foreground
pool
2
Replace the alang-alang thatch on the roof with dark grey clay roof tiles, keeping the same roof shape and pitch.material · roof, large
tiles
3
Change the four heavy ironwood posts and the beams above them from dark stained timber to pale natural honey-coloured teak.material · structure
teak
4
Add a small square open bale bengong pavilion with its own thatched roof at the far end of the pool.add · mid-ground
bale
5
Remove the frangipani tree at the left of the pool, leaving open lawn in its place.remove · left edge
no-frangi
6
Replace the mown lawn between the pool and the steps with white pebble ground and large flat stepping stones.texture · foreground, whole
pebbles
7
Enclose the open-sided central living veranda with full-height sliding glass panels set between the timber posts.form · centre of facade
glass
8
Add a separate single-storey guest building to the right of the villa, in the same timber and thatch, linked to the main house by a low covered walkway.add · composition
second
9
Add a continuous narrow glazed rooflight strip running along the ridge of the main joglo roof.detail · revisits step 2
mahkota
10
Add a narrow shallow stone water channel running from the edge of the pool to a carved stone spout in the garden wall.fine geom · revisits step 1
channel
The keep-clause, appended to every edit

Keep everything else in the image identical — the same building, the same materials, the same objects in the same places.

Do not change the camera in any way: no zooming in or out, no cropping or re-framing, no change of focal length, viewpoint, height or angle, no shifting or rescaling of the scene. The horizon line, the edges of the frame and the position and size of every existing element within the frame must stay exactly where they are. If something new is added, fit it into the existing view rather than pulling the camera back to make room.

Keep exactly the same soft flat overcast midday daylight: do not change the time of day, the weather, the sky, the direction of the light, the shadows or the colour temperature.

The seed prompt

A photorealistic architectural photograph of a large contemporary villa in Bali, Indonesia — a modern architect's reinterpretation of the Javanese joglo, not a traditional house. The joglo is present only as structure and silhouette: a steep two-tiered pyramidal roof in alang-alang thatch with deep clean overhanging eaves, carried on four heavy dark-stained ironwood soko guru posts and plain squared beams. Every surface is flat, planed and unornamented — no carving, no traditional motifs, no decorative fretwork anywhere.

A substantial two-storey private house about twenty-five metres wide, not a garden pavilion. Under the joglo roof at the centre is a deep open-sided living veranda on a honed dark volcanic stone plinth, with three wide stone steps down to the garden. Two long lower wings extend left and right under simple pitched thatch roofs: enclosed rooms clad in flush vertical dark-stained timber boards, with tall frameless floor-to-ceiling glazing and slim dark metal frames. Upstairs bedrooms open onto a continuous first-floor balcony with a plain horizontal timber slat railing running the width of the facade.

A rectangular swimming pool in the foreground, parallel to the facade, edged in dark volcanic stone. Mown lawn between the pool and the steps. A frangipani tree at the left of the pool. A coconut palm grove behind the villa on the right, jungle ridge on the horizon.

Soft flat overcast midday daylight, no direct sun, no strong shadows, neutral white balance. A wide shot photographed at eye level in three-quarter view from the garden side. The entire villa is inside the frame with clearmargin on both sides: the far end of the left wing and the far end of the right wing are both fully visible, and nothing is cut off by the edges of the picture. The whole length of the pool is visible too. No people, no text.

Questions

Which image model degrades least over repeated edits?

In this run, FLUX.3 Image. Over ten sequential architectural edits it ended with the highest edge overlap with its own seed (0.24 against 0.06 to 0.10), the lowest drift (27.3 against 35.5 to 59.6) and the smallest change per edit. It was also the slowest, at 69.7 s median per edit. One run per model, so treat close figures as unresolved.

What does an image model do to a picture over ten edits?

Three different things in this run. MAI-Image-2.6 Flash re-shot the scene from the first edit and kept moving the camera. Seedream 5.0 Flash and GPT Image 2.5 Flare kept the frame but added texture and shifted the sky (sky shift 81.0 and 21.4). FLUX.3 Image changed least (sky shift 4.3) but still did not hold the camera exactly.

Can you stop a model from moving the camera by telling it not to?

Not in this run. Every edit carried a written lock on zoom, crop, re-framing, focal length, viewpoint, height and angle, plus an instruction to fit additions into the existing view. In the frames, all four models still moved the camera over ten edits, by very different amounts; camera movement was not measured separately.

How much did ten edits cost with each model?

GPT Image 2.5 Flare $0.20, FLUX.3 Image $0.24, Seedream 5.0 Flash $0.18, MAI-Image-2.6 Flash $0.23, as reported by the endpoint per request, for ten edits and without the seed. FLUX.3 steps 1 to 8 carry the flat per-image price its other calls carried, because the run was interrupted before the endpoint's own figures were saved.

How can I reproduce this benchmark?

The seed prompt, the ten edits and the keep-clause are published verbatim, with the harness, the seed script, the four seed images, the metrics script and the untouched raw output. You need your own OpenRouter key. The run date and the exact model identifiers and parameters are recorded above, because models change under the same name.

More runs

Sequential editing: what eight edits do to one imageThree editing engines, eight edits in a row.
Three image models, one set of briefsSpeed, brief adherence and aspect ratio.