Skip to main content

Black Forest LabsImage model

FLUX 3 Image

FLUX 3 Image is rank 3 of 108 in GraduallyBench Image v2, our own benchmark.

GraduallyBench Image v2Rank 3 of 108

88our own benchmark0 to 100

Arena ratingnot rated

n/aUser voteshigher is better

Priceno price

n/aper imagelower is better

Popularityno data yet

n/ano search or usage data

Compare with another model

Summary

What is distinctive about FLUX 3 Image is native 2K and 4K output plus references and bounding boxes for arranging objects. Black Forest Labs therefore targets controlled image generation and editing where positions and relationships in the layout matter.

Access runs through the Black Forest Labs API and Playground. Commercial weights are offered on request, while publicly downloadable open weights are not evidenced. FLUX 3 Image remains a proprietary API model.

Image samples from GraduallyBench Image v2

Our own benchmarkThese images come from GraduallyBench Image v2, our own benchmark: every model gets the same tasks with the same prompt, and the first image counts, without selection or retouching.

Character and detailOpen Character and detail at original size

Character and detail

This sample tests Anatomy, clothing, and spatial consistency.

Prompt and settings

Create a square cinematic full-body portrait of a fictional bicycle courier waiting under a transparent umbrella at a rainy tram stop in Hamburg at blue hour. She wears a mustard raincoat, navy trousers, red sneakers, a silver helmet, and carries a teal messenger bag. Keep both hands visible, render the bicycle correctly, use realistic wet-street reflections, and include no readable brand names.

Generated through openrouter with model ID black-forest-labs/flux-3-image. Archived settings were aspect ratio 1:1, resolution 1K, n=1, seed=provider random, actual cost $0.048000.

Open source
InfographicOpen Infographic at original size

Infographic

This sample tests Numbers, labels, and visual organization.

Prompt and settings

Create a square German infographic titled SO FUNKTIONIERT PHOTOSYNTHESE. Show exactly four numbered steps in this order: 1 LICHT, 2 WASSER, 3 CO₂, 4 ZUCKER + SAUERSTOFF. Use a clean editorial science style with one plant cross-section, simple arrows, high contrast, legible labels, and no other text.

Generated through openrouter with model ID black-forest-labs/flux-3-image. Archived settings were aspect ratio 1:1, resolution 1K, n=1, seed=provider random, actual cost $0.048000.

Open source
Product photographyOpen Product photography at original size

Product photography

This sample tests Materials, reflections, and fine details.

Prompt and settings

Create a square premium product photograph of a brushed titanium wristwatch standing upright on dark green marble. The watch has a cream dial, thin black hands set to 10:10, twelve distinct hour markers, a realistic crown, and a dark brown leather strap. Soft window light from the left, controlled reflections, shallow depth of field, no text, no logo, no extra objects.

Generated through openrouter with model ID black-forest-labs/flux-3-image. Archived settings were aspect ratio 1:1, resolution 1K, n=1, seed=provider random, actual cost $0.048000.

Open source
Typography and layoutOpen Typography and layout at original size

Typography and layout

This sample tests Legible text, hierarchy, and composition.

Prompt and settings

Create a square editorial poster for an imaginary night train called MONDFALTER. Show the exact German headline MONDFALTER and the exact subline BERLIN NACH LISSABON. Use a restrained midnight-blue and warm-cream palette, one stylized moth, strong Swiss-grid typography, generous negative space, and no additional words or logos.

Generated through openrouter with model ID black-forest-labs/flux-3-image. Archived settings were aspect ratio 1:1, resolution 1K, n=1, seed=provider random, actual cost $0.048000.

Open source

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 265 published benchmark results

Page 1 of 23

Published benchmarks

GraduallyBench Image v2: Typography and layout

The 0-100 value is the estimated chance to beat an average model of the test cohort in model-blind pairwise comparisons. One archived first output contributes per model, judged by two automated judge models. This remains an n = 1 sample.

FLUX 3 Image95.8 points
Muse Image95.3 points
Nano Banana Pro93.2 points
Nano Banana 2.192.5 points
Nano Banana 292 points
GPT Image 288.8 points
Wan 2.7 Image87.1 points
Ideogram 4.084.8 points
FLUX.2 Flex80.7 points
Seedream 4.578 points
MAI-Image-2.577.8 points
Seedream 5.0 Pro76.5 points
FLUX.2 Max76.2 points
Krea 2 Large73.8 points
Recraft V4.170.6 points
Krea 370.4 points
Bria FIBO 269.8 points
GPT Image 1.568 points
LLaDA-Image63.9 points
MAI-Image-2.663.6 points
Luma Uni-1.162.1 points
FLUX.2 Pro61.5 points
Luma Uni-1.1 Max59.1 points
Reve 2.158.3 points
MOSS-Image-Base57.6 points
FLUX.2 Dev56.4 points
GPT Image 1 Mini53.1 points
Ideogram 4.546.5 points
GLM-Image44.1 points
Nucleus-Image42.9 points
FLUX.2 Klein 9B42.8 points
Seedream 4.041.8 points
FLUX1.1 [pro]37.9 points
Z-Image Turbo37.9 points
Qwen Image 3.036.7 points
Kling Image 3.034.9 points
Z-Image Base34.5 points
HiDream-O1-Image34.4 points
LongCat-Image34.2 points
Chroma1-HD29.7 points
Bria FIBO Lite29.4 points
MOSS-Image-SFT28.1 points
MiniMax Image-0124.5 points
Chroma1-Flash23.2 points
HunyuanImage 3.021.9 points
Midjourney V8.121.7 points
BAGEL-7B-MoT21.4 points
Vidu Q2 Image20.1 points
OmniGen217.8 points
FLUX.1 [dev]12.4 points
AuraFlow v0.310.9 points
SDXL Turbo3.8 points

higher is betterAs of November 7, 2026Scale: 0 points - 100 points

Source: Gradually-Bildtest
gradually.ai
Methodology and context
Metric
pairwise-score-100
Scale
0-100
Participants
108
Sample
1
Methodology and context
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source: Gradually-Bildtest

GraduallyBench Image v2: Product photography

The 0-100 value is the estimated chance to beat an average model of the test cohort in model-blind pairwise comparisons. One archived first output contributes per model, judged by two automated judge models. This remains an n = 1 sample.

Qwen Image 3.093.3 points
Wan 2.7 Image93.2 points
Krea 2 Large89.1 points
Muse Image88.3 points
MAI-Image-2.687.2 points
Nano Banana 2.185.2 points
Bria FIBO 284.8 points
FLUX.2 Max81.3 points
GPT Image 277.9 points
Seedream 5.0 Pro76.6 points
Nano Banana Pro71.4 points
FLUX.2 Flex71.1 points
Ideogram 4.570.7 points
FLUX 3 Image70.1 points
FLUX.2 Klein 9B69.7 points
Nucleus-Image69.4 points
Luma Uni-1.1 Max68.5 points
MAI-Image-2.567 points
Reve 2.166.4 points
Luma Uni-1.164.4 points
Krea 364.1 points
Recraft V4.162.2 points
Midjourney V8.160.3 points
FLUX.2 Klein 4B58.6 points
GPT Image 1.558.2 points
FLUX.2 Dev57.9 points
MOSS-Image-Base57.8 points
MOSS-Image-SFT56.9 points
FLUX.2 Pro53.6 points
BAGEL-7B-MoT51.3 points
GPT Image 1 Mini47.9 points
Kling Image 3.046.6 points
Nano Banana 245.3 points
Bria FIBO Lite45.1 points
Seedream 4.544.7 points
Seedream 4.043.8 points
Ideogram 4.041.5 points
FLUX.1 [dev]40.8 points
LLaDA-Image39 points
HiDream-O1-Image30.2 points
Z-Image Turbo30.2 points
Vidu Q2 Image27.5 points
HunyuanImage 3.026.4 points
LongCat-Image26.3 points
Z-Image Base24.6 points
OmniGen223.6 points
Chroma1-Flash23.5 points
Lumina-Image 2.017.5 points
FLUX1.1 [pro]17.4 points
MiniMax Image-0117.2 points
GLM-Image12 points
AuraFlow v0.311.6 points
Chroma1-HD8.7 points
SDXL Turbo1.9 points

higher is betterAs of November 7, 2026Scale: 0 points - 100 points

Source: Gradually-Bildtest
gradually.ai
Methodology and context
Metric
pairwise-score-100
Scale
0-100
Participants
108
Sample
1
Methodology and context
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source: Gradually-Bildtest

GraduallyBench Image v2: Character and detail

The 0-100 value is the estimated chance to beat an average model of the test cohort in model-blind pairwise comparisons. One archived first output contributes per model, judged by two automated judge models. This remains an n = 1 sample.

FLUX 3 Image95.7 points
GPT Image 1.591.3 points
Ideogram 4.091.2 points
Muse Image88.4 points
Bria FIBO 286 points
FLUX.2 Pro84.1 points
GPT Image 284.1 points
Krea 382.8 points
Wan 2.7 Image82.2 points
FLUX.2 Dev79.9 points
MAI-Image-2.578.7 points
Seedream 5.0 Pro78.1 points
Nano Banana 2.176.6 points
MOSS-Image-Base68.5 points
Qwen Image 3.067.9 points
Qwen-Image-2.167.5 points
FLUX.2 Flex65 points
FLUX.2 Max61 points
Vidu Q2 Image61 points
Z-Image Base59.6 points
Luma Uni-1.1 Max59.4 points
Krea 2 Large58.8 points
Seedream 4.053.6 points
MAI-Image-2.653 points
Bria FIBO Lite51.9 points
FLUX.2 Klein 9B51.6 points
Nucleus-Image50.7 points
HiDream-O1-Image50.4 points
Luma Uni-1.149.7 points
Nano Banana 248.9 points
FLUX.2 Klein 4B47.9 points
HunyuanImage 3.047.5 points
FLUX.1 [dev]45.3 points
Reve 2.141.3 points
Z-Image Turbo40.5 points
Ideogram 4.538.9 points
Kling Image 3.038.4 points
MOSS-Image-SFT33.5 points
LLaDA-Image32.9 points
Chroma1-Flash29.8 points
LongCat-Image29.1 points
Recraft V4.128.1 points
Seedream 4.527 points
Chroma1-HD24.3 points
GLM-Image24.3 points
MiniMax Image-0123.3 points
FLUX1.1 [pro]22.6 points
BAGEL-7B-MoT17.2 points
Lumina-Image 2.012.5 points
AuraFlow v0.312.2 points
GPT Image 1 Mini12.2 points
OmniGen210.1 points
SDXL Turbo4.7 points

higher is betterAs of November 7, 2026Scale: 0 points - 100 points

Source: Gradually-Bildtest
gradually.ai
Methodology and context
Metric
pairwise-score-100
Scale
0-100
Participants
108
Sample
1
Methodology and context
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source: Gradually-Bildtest

GraduallyBench Image v2: Infographic

The 0-100 value is the estimated chance to beat an average model of the test cohort in model-blind pairwise comparisons. One archived first output contributes per model, judged by two automated judge models. This remains an n = 1 sample.

Seedream 5.0 Pro96.9 points
FLUX 3 Image91.7 points
Qwen Image 3.091.1 points
Nano Banana 2.189.9 points
Reve 2.189.5 points
Muse Image88.4 points
Ideogram 4.085.1 points
GPT Image 283.3 points
Luma Uni-1.182.3 points
Seedream 4.081.7 points
Luma Uni-1.1 Max81.3 points
Ideogram 4.574.9 points
Krea 2 Large74 points
Krea 371.5 points
Qwen-Image-2.169.5 points
Nano Banana Pro66.6 points
GPT Image 1.566.2 points
MOSS-Image-SFT63.3 points
Bria FIBO 263.1 points
MAI-Image-2.562.6 points
Wan 2.7 Image62.4 points
MAI-Image-2.661 points
FLUX.2 Dev58.9 points
Recraft V4.158 points
Nano Banana 256.6 points
FLUX.2 Pro49.1 points
FLUX.2 Max48.8 points
MOSS-Image-Base47.3 points
Seedream 4.546.5 points
FLUX.2 Flex45.9 points
FLUX.2 Klein 9B44.8 points
Kling Image 3.044.8 points
Midjourney V8.143.1 points
Z-Image Turbo37.3 points
LongCat-Image35.6 points
MiniMax Image-0134.4 points
GLM-Image34 points
GPT Image 1 Mini30.7 points
HunyuanImage 3.029.9 points
Vidu Q2 Image29 points
LLaDA-Image28.9 points
FLUX.1 [dev]28.1 points
Z-Image Base24.5 points
Chroma1-HD23.1 points
FLUX1.1 [pro]22.7 points
Chroma1-Flash21.9 points
FLUX.2 Klein 4B21.4 points
Bria FIBO Lite15.8 points
Lumina-Image 2.013.1 points
Nucleus-Image12.8 points
BAGEL-7B-MoT10.9 points
OmniGen210.2 points
AuraFlow v0.38.8 points
SDXL Turbo7.7 points

higher is betterAs of November 7, 2026Scale: 0 points - 100 points

Source: Gradually-Bildtest
gradually.ai
Methodology and context
Metric
pairwise-score-100
Scale
0-100
Participants
108
Sample
1
Methodology and context
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source: Gradually-Bildtest

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

AA Image Arena Elo (Show measurement, test conditions, and source)
Source value
1,114
Metric
AA Image Arena Elo
Unit
elo
Benchmark version
AA-Image-T2I v2.1
Category
text-to-image
Direction
Higher is better
Sample
5672
Lower confidence bound
1101
Upper confidence bound
1127
Source type
independent-evaluator
Evaluator
Artificial Analysis AA-Image-T2I v2.1
Status
active
Retrieved at
2026-10-15
Methodology
Independent evaluation. Overall text-to-image; exact published row FLUX 3 Image.
Source Artificial Analysis AA-Image-T2I v2.1
Rapidata General Image Benchmark Elo (Show measurement, test conditions, and source)
Source value
1,122.8456211021
Metric
Bradley-Terry Elo rating from pairwise human preference votes
Unit
elo
Benchmark version
Text-to-Image v1, General image benchmark; snapshot last updated 2026-10-17
Category
text-to-image
Direction
Higher is better
Lower confidence bound
1113.7384499451498
Upper confidence bound
1131.952792259147
Source type
independent-evaluator
Evaluator
Rapidata benchmark.ai
Status
superseded
Retrieved at
2026-10-17
Methodology
Independent evaluation. Overall General image text-to-image preference row, exact displayed label FLUX 3 Image (source row id par_1WlGTIUVtmsmSV; benchmark id bmk_1Pet6iHzbf6RQu). Fixed prompt set across multiple categories; human annotators compare outputs for the same prompt; Bradley-Terry fit rescaled to Elo; published 95% bootstrap interval. Each match is decided by a majority of at least five judgements from different annotators. The source does not publish an immutable serving revision, API/UI surface, resolution, seed or generation controls. The raw totalMatches field is fractional and is not treated as an exact integer N.
Context
The result is attached only to the exact displayed FLUX 3 Image label. Serving revision and generation mode are not disclosed; no FLUX-family sibling or separate image-editing score is transferred.
Source Rapidata benchmark.ai
BudgetPixel AI Image Arena: overall blind-preference rating (Show measurement, test conditions, and source)
Source value
1,507.22
Metric
Bradley-Terry rating from blind pairwise image preferences; 1500 is the platform average
Unit
elo
Benchmark version
BudgetPixel AI Image Arena, Overall, snapshot computed 2026-10-17T06:15:14.668833Z
Category
text-to-image
Direction
Higher is better
Sample
1554
Lower confidence bound
1490.03
Upper confidence bound
1523.47
Source type
independent-evaluator
Evaluator
BudgetPixel AI Image Arena overall leaderboard
Status
superseded
Retrieved at
2026-10-17
Methodology
Independent evaluation. Overall BudgetPixel AI Image Arena result. Exact row id and display name are stated per observation; native JSON snapshot computed at 2026-10-17T06:15:14.668833Z. BudgetPixel says visitors make blind pairwise choices between images from the same prompt; votes are refit nightly with Bradley-Terry on an Elo scale where 1500 is average and the interval is a 95% bootstrap range. Its page says arena generations use the same prompt, default settings, and resolution. The source's votes field counts comparisons, not unique users or prompts. No immutable underlying provider checkpoint or API revision is published. Exact source row id flux-3-image, display name “FLUX 3 Image”, maker Black Forest Labs; rank 10; votes=1554. The native public JSON contains the exact decimal score and interval endpoints; the HTML table rounds those values.
Context
This rating belongs only to BudgetPixel's exact named arena route and this computed snapshot. The source does not identify a pinned provider API/checkpoint beneath that route; do not transfer the score to another arena or generation setting.
Source BudgetPixel AI Image Arena overall leaderboard
HELDER FLUX 3 Image review, Average IoU across prescribed boxes (Show measurement, test conditions, and source)
Source value
0.89
Metric
Mean intersection over union across 16 object boxes, 0–1; higher is better
Unit
ratio
Category
image composition, editing, or text rendering
Direction
Higher is better
Sample
16
Source type
community-reported
Evaluator
HELDER Staff, FLUX 3 Image practical image-editing and composition tests
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Tests were run through fal.ai on 1 October 2026. Four scenes had four prescribed object locations each; FLUX 3 received JSON bounding boxes while competing models received percentage-based text instructions. The article states one attempt per model.
Context
The bounding-box inputs differ by model; no immutable API revision or confidence interval is published. SSIM is one five-edit sequence, not a repeated-sample estimate.
Source HELDER Staff, FLUX 3 Image practical image-editing and composition tests
HELDER FLUX 3 Image review, Objects placed in assigned boxes (Show measurement, test conditions, and source)
Source value
16
Metric
Count of objects within their specified boxes, out of 16
Unit
count
Category
image composition, editing, or text rendering
Direction
Higher is better
Sample
16
Source type
community-reported
Evaluator
HELDER Staff, FLUX 3 Image practical image-editing and composition tests
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Tests were run through fal.ai on 1 October 2026. Four scenes had four prescribed object locations each; FLUX 3 received JSON bounding boxes while competing models received percentage-based text instructions. The article states one attempt per model.
Context
The bounding-box inputs differ by model; no immutable API revision or confidence interval is published. SSIM is one five-edit sequence, not a repeated-sample estimate.
Source HELDER Staff, FLUX 3 Image practical image-editing and composition tests
HELDER FLUX 3 Image review, Untouched-region SSIM after five edits (Show measurement, test conditions, and source)
Source value
0.995
Metric
Structural similarity of unchanged image regions after a five-edit sequence, 0–1
Unit
ratio
Category
image composition, editing, or text rendering
Direction
Higher is better
Source type
community-reported
Evaluator
HELDER Staff, FLUX 3 Image practical image-editing and composition tests
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Tests were run through fal.ai on 1 October 2026. Four scenes had four prescribed object locations each; FLUX 3 received JSON bounding boxes while competing models received percentage-based text instructions. The article states one attempt per model.
Context
The bounding-box inputs differ by model; no immutable API revision or confidence interval is published. SSIM is one five-edit sequence, not a repeated-sample estimate.
Source HELDER Staff, FLUX 3 Image practical image-editing and composition tests
HELDER FLUX 3 Image review, Polish sign text exactly reproduced (Show measurement, test conditions, and source)
Source value
6
Metric
Count of exactly reproduced sign texts, out of 8
Unit
count
Category
image composition, editing, or text rendering
Direction
Higher is better
Sample
8
Source type
community-reported
Evaluator
HELDER Staff, FLUX 3 Image practical image-editing and composition tests
Status
active
Retrieved at
2026-10-17
Methodology
Community-reported result. Tests were run through fal.ai on 1 October 2026. Four scenes had four prescribed object locations each; FLUX 3 received JSON bounding boxes while competing models received percentage-based text instructions. The article states one attempt per model.
Context
The bounding-box inputs differ by model; no immutable API revision or confidence interval is published. SSIM is one five-edit sequence, not a repeated-sample estimate.
Source HELDER Staff, FLUX 3 Image practical image-editing and composition tests
BudgetPixel AI Image Arena: Product & commercial category rating (Show measurement, test conditions, and source)
Source value
1,563
Metric
Category-specific Bradley-Terry Elo rating from blind head-to-head image preferences in Product & commercial; the platform states that 1500 is average.
Unit
elo
Category
Product & commercial
Direction
Higher is better
Lower confidence bound
1516
Upper confidence bound
1621
Source type
independent-evaluator
Evaluator
BudgetPixel AI Image Arena, FLUX 3 Image category results
Status
superseded
Retrieved at
2026-10-21
Methodology
Independent evaluation. BudgetPixel AI Image Arena, Product & commercial category. Blind head-to-head votes compare images generated from the identical verbatim prompt at each model's default settings and resolution, with a fixed seed where supported. Votes are refit nightly with Bradley-Terry Elo (1500 average); ties count half. The model page snapshot is as of October 16, 2026; the source reports 145 pairwise-comparison votes involving FLUX 3 Image in this category. This is not a count of unique tasks, prompts, or raters, nor an effective N.
Context
This is a BudgetPixel-specific human-preference rating from a rolling text-to-image prompt suite, not an absolute or independently reproduced image-quality score. The arena updates nightly. The printed vote count is model/category vote exposure, not a count of distinct prompts. The page does not disclose an immutable served checkpoint or exact generation controls for this model route.
Source BudgetPixel AI Image Arena, FLUX 3 Image category results

Recorded measurement details

GraduallyBench Image v2: Typography and layout (Show measurement, test conditions, and source)
Source value
95.8
Metric
pairwise-score-100
Unit
pairwise score
Direction
Higher is better
Source rank
2
Sample
1
Source type
gradually-sample-test
Evaluator
Gemini 4 Argon (High) und GPT-6 Astra
Retrieved at
2026-11-07
Methodology
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source Gradually-Bildtest
GraduallyBench Image v2: Product photography (Show measurement, test conditions, and source)
Source value
70.1
Metric
pairwise-score-100
Unit
pairwise score
Direction
Higher is better
Source rank
29
Sample
1
Source type
gradually-sample-test
Evaluator
Gemini 4 Argon (High) und GPT-6 Astra
Retrieved at
2026-11-07
Methodology
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source Gradually-Bildtest
GraduallyBench Image v2: Character and detail (Show measurement, test conditions, and source)
Source value
95.7
Metric
pairwise-score-100
Unit
pairwise score
Direction
Higher is better
Source rank
1
Sample
1
Source type
gradually-sample-test
Evaluator
Gemini 4 Argon (High) und GPT-6 Astra
Retrieved at
2026-11-07
Methodology
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source Gradually-Bildtest
GraduallyBench Image v2: Infographic (Show measurement, test conditions, and source)
Source value
91.7
Metric
pairwise-score-100
Unit
pairwise score
Direction
Higher is better
Source rank
4
Sample
1
Source type
gradually-sample-test
Evaluator
Gemini 4 Argon (High) und GPT-6 Astra
Retrieved at
2026-11-07
Methodology
One archived image per model, using the first output of the accepted test attempt. 648 image pairs of this task are compared model-blind in both orders by two judge models from different vendors (Gemini 4 Argon (High) and GPT-6 Astra). The 0-100 value is the estimated chance to beat an average model of the test cohort; the range is the 95 % interval. This remains an n = 1 sample, and judge models can have visual preferences of their own.
Source Gradually-Bildtest

Page 1 of 23

All specifications

Architecture

Model class
Text-to-image and image-editing model
Source: Model class

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Subject to limitations. Black Forest Labs describes an optional invisible soft-binding watermark for its image output pipeline. Whether a particular file contains it depends on the output path.
Source: AI watermarking
Watermarking technique
Not known
Source: Watermarking technique
Provenance metadata
C2PA Content Credentials for final images through the BFL API, Playground and BFL-controlled output pipelines. BFL makes no such guarantee for local execution or third-party providers.
Source: Provenance metadata
Labelling reviewed
Nov 5, 2026
Source: Labelling reviewed

Licence and access

Access
API, Web
Source: Access
Weights
Commercial weights on request, no public open weights
Source: Weights
Shutdown
None announced
no source

API

API model ID
flux-3-image
Source: API model ID

Further details

Access
Prompt-to-image API and BFL Playground; no public weights are listed
Source: Access
Release context
Early-access announcement on July 23, 2026; general API availability date is not stated
Source: Release context
Resolution
Native 2K and 4K output
Source: Resolution
Layout control
Layout control with references and bounding boxes
Source: Layout control

Prices

Choose a price view
OpenRouter (black-forest-labs/flux-3-image)
$0.048 per image (estimated)Source
fal.ai (blackforestlabs/flux-3/text-to-image)
$0.048 per image (estimated)Source

No Black Forest Labs subscription access recorded. Through other providers from $9 a month (Krea Basic, with limited credits). 20 plans in Germany.

Filter plans
PlanMonthly list priceNote
Krea Basic$9with limited credit
Leonardo.Ai Essential$12
Runway Standard$15with limited credit
Leonardo.Ai Premium$30
Krea Pro$35with limited credit
Runway Pro$35with limited credit
Genspark for Business Team$40with limited credit, for teams, per seat
Leonardo.Ai Ultimate$60
Higgsfield Plus€59with limited credit
Runway Team$69with limited credit, for teams, per seat
Higgsfield Team€79with limited credit, for teams, per seat
Runway Max$95with limited credit
Krea Max$105with limited credit
Higgsfield Ultra€129with limited credit
Higgsfield Scale€169with limited credit, for teams, per seat
Krea Business$200with limited credit, for teams
Genspark for Business Enterprisen/awith limited credit, for enterprises, per seat
Higgsfield Enterprisen/afor enterprises
Krea Enterprisen/afor enterprises
Runway Enterprisen/afor enterprises

All plans in the AI subscription comparison

Frequently asked questions about FLUX 3 Image

How good is FLUX 3 Image?

Rank 3 of 108 in GraduallyBench Image v2, our own benchmark.

GraduallyBench Image v2 from Nov 7, 2026

GraduallyBench Image v2our own benchmark, 0 to 100 · higher is better
Muse Image90
Grok Imagine Image 2.090
FLUX 3 ImageRank 3 of 10888
Nano Banana 2.186
GPT Image 2.5 Sunburst86
FLUX 3 Image Klein85
GPT Image 2.5 Flare85

What is FLUX 3 Image good at?

Best at Typography and layout (96 out of 100), weakest at Product photography (70).

GraduallyBench Image v2, one image per task

  1. FLUX 3 Image: Typography and layoutTypography and layoutLegible text, hierarchy, and composition96out of 100Rank 2 of 108 in this task
  2. FLUX 3 Image: Product photographyProduct photographyMaterials, reflections, and fine details70out of 100Rank 29 of 108 in this task
  3. FLUX 3 Image: Character and detailCharacter and detailAnatomy, clothing, and spatial consistency96out of 100Rank 1 of 108 in this task
  4. FLUX 3 Image: InfographicInfographicNumbers, labels, and visual organization92out of 100Rank 4 of 108 in this task

Where can I use FLUX 3 Image?

No Black Forest Labs subscription access recorded. Including other providers that makes 20 plans in Germany.

Subscription comparison, prices in Germany incl. VAT

At Black Forest Labs

No Black Forest Labs subscription access recorded.

Through other providers

20 plans from 5 providers, from $9 a month (Krea Basic), with limited credits.

Sources and data date

Show all 36 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 32 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

quality-benchmark

Original wording

No exact FLUX 3 Image quality score, vote cohort, or confidence interval was published. The separately measured FLUX 3 video row is excluded.

Not published

exact-serving-and-generation-configuration

Note

The evaluator does not publish an immutable serving revision, API/UI surface, image resolution, seed or generation controls for this row.

Not published

immutable-serving-revision

Note

Neither evaluator publishes an immutable provider serving/checkpoint revision or full generation controls for its named route; the recorded cells are limited to each exact public leaderboard label and snapshot.

Additional sourceBlack Forest Labs (retrieved October 3, 2026; October 4, 2026; October 15, 2026; November 1, 2026) · Black Forest Labs FLUX 3 Image · Editorial description reviewed October 3, 2026
Additional sourceBlack Forest Labs (retrieved November 1, 2026)
Additional sourceOpenRouter
Additional sourcefal.ai
Additional sourceGradually-Bildtest (retrieved November 7, 2026)
Additional sourceArtificial Analysis AA-Image-T2I v2.1 (retrieved October 15, 2026)
Additional sourceRapidata benchmark.ai (retrieved October 17, 2026; October 28, 2026)
Additional sourceBudgetPixel AI Image Arena overall leaderboard (retrieved October 17, 2026; October 28, 2026; November 17, 2026)
Additional sourceHELDER Staff, FLUX 3 Image practical image-editing and composition tests (retrieved October 17, 2026)
Additional sourceBudgetPixel AI Image Arena, FLUX 3 Image category results (retrieved October 21, 2026)
Additional sourceArtificial Analysis (retrieved October 23, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Architecture & interiors) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Concept art & fantasy) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Glamour) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Illustration & anime) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Landscape & nature) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Overall) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Photorealism) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Portrait & fashion) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Product & commercial) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Prompt adherence) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceBudgetPixel AI Image Arena public leaderboard CSV (Typography) (retrieved October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 3, 2026; November 4, 2026; November 5, 2026; November 6, 2026; November 7, 2026; November 8, 2026; November 9, 2026; November 10, 2026; November 12, 2026; November 13, 2026; November 14, 2026; November 15, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceModel metadata (retrieved October 17, 2026; November 19, 2026)
Additional sourceModel metadata (retrieved October 28, 2026; November 13, 2026; November 14, 2026; November 16, 2026; November 17, 2026; November 19, 2026)
Additional sourceModel metadata (retrieved October 28, 2026)
Additional sourceModel metadata (retrieved November 19, 2026)12345678910
Additional sourceBlack Forest Labs: Content Credentials (retrieved November 5, 2026)