Skip to content
trending

Benchmarks

FLUX 3 Image: 5 Edits in One Prompt vs 6 Other Models

All seven models made five changes from one prompt, but only FLUX 3 Image left the rest of the photo untouched. SSIM 0.995 vs Grok's 0.708.

By 8 min read
FLUX 3 Image: 5 Edits in One Prompt vs 6 Other Models

👁 279 reads

All seven models executed five changes from a single prompt, but only FLUX 3 Image left the rest of the photo practically untouched. In four areas that should have remained unchanged, similarity to the original (SSIM) was 0.995 out of 1. Right behind it was Ideogram 4.5 (0.989), followed by Seedream 5 Pro (0.961) and Nano Banana Pro (0.942). GPT Image 2.5 (0.870) and Grok Imagine 2 (0.708) repainted the entire room. The price for this precision: 204 seconds of waiting and a 4 Mpx limit on input images.

Black Forest Labs states that FLUX 3 Image handles multiple edits in one round, and pixels outside the edited areas remain “bit-for-bit identical.” This is an important promise for anyone editing product photos or interior visualizations: you change the sofa color, and the rest of the photo should stay exactly the same, including the parquet texture and molding.

I tested that promise on October 1, 2026, launch day, with one photo of a living room and one prompt containing five changes, then gave the same job to six other models that edit images from text.

FLUX Hub - Black Forest Labs models, tests and prompts: FLUX 3 Image and FLUX.2.

One photo, one prompt, one attempt per model

  • Date: October 1, 2026, FLUX 3 Image launch day.
  • Access: all models through edit endpoints on fal.ai.
  • Base image: apartment living room generated by Nano Banana Pro, 2528 × 1696 px. Every model received the same image.
  • Exception: FLUX 3 on fal.ai accepts input images with a maximum area of 4 Mpx. The first call with the full file (4.29 Mpx) returned an error “Image dimensions are too large”, so FLUX received the same base scaled down to 2400 × 1610 px.
  • Single attempt: one result per model, no randomization or cherry-picking better versions.
  • Same prompt for all, in English.
  • Measurement: I checked each change myself at high magnification and measured the unchanged areas with SSIM in four regions: window with curtains, molding under the ceiling, right wall, and parquet floor. SSIM 1 means an image identical to the original.

The prompt: five changes at once

Make exactly these five changes and keep everything else in the photo identical: 1) change the sofa upholstery from dark green to rust terracotta velvet, same shape and cushions; 2) replace the abstract painting with a black and white photograph of the Palace of Culture and Science in Warsaw, keep the same gold frame and position; 3) remove the ceramic vase from the coffee table; 4) change the floor lamp shade from cream to matte black; 5) add a small sleeping brown dachshund lying on the striped rug.

The changes are deliberately varied:

  1. Sofa color - material change without altering the shape.
  2. Wall painting - content replacement while keeping the frame and position.
  3. Vase removal - the model must redraw the table surface underneath.
  4. Lamp shade - small change in one element of a larger object.
  5. Dachshund on rug - adding a new object with shadow and perspective.

FLUX 3 Image made all five changes and left the room alone

Apartment living room before editing: green sofa, abstract painting, vase on table, cream lamp shade
Before · base
Living room after FLUX 3 Image edit: terracotta sofa, Palace of Culture photo, black lamp shade, dachshund on rug
After · FLUX 3 Image · SSIM 0.995

All five changes are in place. The sofa has rust velvet and the same cushions, the gold frame holds a black and white photo of the Palace of Culture, the vase is gone and the table surface beneath it is clean. The lamp shade is black, the lamp base remained brass, and the dachshund sleeps on the rug to the right of the table with a shadow beneath it.

I checked the “bit-for-bit” promise separately by comparing the result with the 2400 × 1610 px file that FLUX received as input. However, fal.ai returned an image of 2480 × 1664 px, a different size than the input, so I had to rescale the pixels and cannot confirm exact identity. The difference in unchanged areas averaged between 0.45 and 1.8 brightness levels out of 255, which is invisible. On the difference map, only the sofa, painting, lamp shade, vase location, and dachshund shine brightly.

Six other models, from Ideogram 4.5 to Grok Imagine 2

Living room after Ideogram 4.5 edit: orange sofa, Palace of Culture photo, dachshund by table
Ideogram 4.5 · SSIM 0.989
Living room after Seedream 5 Pro edit: terracotta sofa, Palace of Culture in frame, dachshund on rug
Seedream 5 Pro · SSIM 0.961
Living room after Nano Banana Pro edit: Palace of Culture photo with white mat, dachshund next to table
Nano Banana Pro · SSIM 0.942
Living room after Qwen Image 3 edit: terracotta sofa, black lamp shade, dachshund on rug
Qwen Image 3 · SSIM 0.889
Living room after GPT Image 2.5 edit: all changes made but entire interior repainted
GPT Image 2.5 Sunburst · SSIM 0.870
Living room after Grok Imagine 2 edit: entire floor lamp repainted black
Grok Imagine 2 · SSIM 0.708

What each model did:

  • Ideogram 4.5: complete set of changes and second-best SSIM score. The sofa came out more orange than rust, but it fits the prompt. The dachshund lies on the rug right next to the table leg.
  • Seedream 5 Pro: five changes, frame preserved, rest of the room slightly changed in textures.
  • Nano Banana Pro: five changes, but added a white mat to the Palace of Culture photo that wasn’t in the frame. Same frame, but the interior of the frame changed.
  • Qwen Image 3: five changes, noticeably altered curtains and parquet, which isn’t visible at first glance but shows on the difference map.
  • GPT Image 2.5 Sunburst: five changes, also with added mat. The model repainted the entire interior: curtains, molding, ficus leaves, and parquet differ from the original in small details.
  • Grok Imagine 2: did everything, but painted the entire lamp black, including the stand and base, not just the shade. It also slightly shifted and rescaled the entire interior, hence the lowest SSIM.

Results: all 5/5, but SSIM from 0.995 down to 0.708

Model Changes Completed Additional, Unordered Changes SSIM of Unchanged Areas Time Price on fal.ai
FLUX 3 Image 5/5 none 0.995 204 s $0.024 (1K) on launch day; $0.048 since October 8
Ideogram 4.5 5/5 none visible 0.989 up to ~3 min* $0.22 (high)
Seedream 5 Pro 5/5 minor textures 0.961 up to ~3 min* $0.135
Nano Banana Pro 5/5 mat in frame 0.942 up to ~3 min* $0.15 (2K)
Qwen Image 3 5/5 curtains, parquet 0.889 up to ~3 min* $0.075 (2K)
GPT Image 2.5 Sunburst 5/5 mat, entire interior repainted 0.870 up to ~3 min* size-dependent (2560 × 1440 high: $0.055)
Grok Imagine 2 5/5 entire lamp black, shifted frame 0.708 up to ~3 min* $0.08 (2K medium) + $0.01 for input image

* I timed only FLUX 3 exactly (second run, after downsizing the base). Other models responded on the first pass, and according to file save times, each returned results within about three minutes of starting.

SSIM in areas that should have remained unchanged (1 = identical)
FLUX 3 Image0.995
Ideogram 4.50.989
Seedream 5 Pro0.961
Nano Banana Pro0.942
Qwen Image 30.889
GPT Image 2.5 Sunburst0.870
Grok Imagine 20.708

What SSIM Shows and What It Doesn’t

SSIM compares brightness, contrast, and structure of two image fragments. The difference between 0.995 and 0.942 looks innocent on the chart, but in practice it’s the boundary between “I see no change” and “the parquet and curtains look slightly different when I toggle between images.” At Grok’s 0.708, you see the difference immediately because the entire room shifted.

However, SSIM says nothing about the quality of the changes themselves. Ideogram has almost as high an SSIM as FLUX, but made the sofa more orange. GPT Image 2.5 has low SSIM, but its dachshund and Palace of Culture photo look very good. That is why I score the two separately.

When precision matters more than looks

  • Product and real estate photos: clients compare “before” and “after” versions pixel by pixel. Here FLUX 3 and Ideogram 4.5 are safe, while GPT Image 2.5 and Grok Imagine 2 may change something no one ordered.
  • Series of variants: five sofa colors in the same interior. A model that repaints the background will give you five slightly different rooms.
  • Creative rework: if you care about a new atmosphere rather than fidelity, repainting everything isn’t a flaw.

FLUX 3’s limitation is input size. A 24 Mpx camera photo must be downsized to 4 Mpx before editing, so “bit-for-bit” applies to that downsized version, not the original.

One prompt beats ten rounds of edits

This test checked five changes in one prompt. That’s a different situation from ten corrections in sequence, where each step can add small distortions. I tested Ideogram 4.5 earlier in exactly that scenario: ten sequential edits of one photo. If you know in advance what you want to change, collect everything in one prompt. Fewer rounds means fewer opportunities to corrupt the rest of the image.

Five rules for a multi-change prompt

  1. Start with a prohibition. “Make exactly these five changes and keep everything else identical” sets the model in edit mode, not repaint mode.
  2. Number the changes. Each in a separate point, with the name of the object the model should find.
  3. State before and after. “From dark green to rust terracotta” is more reliable than just “terracotta”.
  4. Say what stays. “Same gold frame and position”, “same shape and cushions”. Two models still added a mat, so for frames add “no mat”.
  5. Specify the part of the object. Grok understood “lamp shade” as the entire lamp. Adding “keep the brass stand” would probably have helped.

More FLUX 3 tests

FAQ

Does FLUX 3 really leave the rest of the photo unchanged?

In my test, the difference in unchanged areas was invisible (SSIM 0.995, averaging below 2 brightness levels out of 255). I didn’t confirm bit-for-bit identity because fal.ai returned an image in a different size than the input.

What image can you upload for editing in FLUX 3 on fal.ai?

With an area of at most 4 Mpx, for example 2400 × 1610 px. The API rejects larger files with the error “Image dimensions are too large”.

How much does editing in FLUX 3 cost?

On fal.ai a 1K (1 MP) result costs $0.048, at a flat price per image regardless of how many reference images you send; larger outputs cost more. The launch-day price of $0.024 was introductory and ended on October 8, 2026. Editing took 204 seconds in my test.

Kinetyra

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editorial team

Independent AI publication. Models and tools are tested on real tasks, prices are checked against official rate cards, and articles are updated when things change.

read nextFLUX 3 4K Test: Real Resolution vs 6 AI Image ModelsbenchmarksFLUX 3 4K Test: Real Resolution vs 6 AI Image ModelsFLUX 3 Image delivers true 4K with 16.8 Mpx of real detail. We tested it against six models to measure effective resolution, text accuracy, and speed.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages1

Quantslant in numbers 1,457posts in archive37news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›