---
title: "FLUX.2 Klein Blind Test Results: Scores by Use Case"
canonical_url: "https://workroom.everypixel.com/models/flux-2-klein"
score: 23.9
rank: 5
scale: "Anchor comparison of generative image models, v3"
tested: 2026-08-25
---

# FLUX.2 Klein Blind Test Results: Scores by Use Case

**Score:** 23.9. A real photograph is 100.
**Rank:** #5 on the scale
**Author:** Dmitry Shironosov (Head of AI Research, Everypixel Workroom)
**Tested:** 2026-08-25
**Canonical URL:** https://workroom.everypixel.com/models/flux-2-klein

## Independence

Workroom is an AI router: we give access to third-party models, including FLUX.2 Klein, and we are not affiliated with Black Forest Labs. We have no stake in any single vendor winning; the router is only useful if it sends each brief to the right model. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0 and linked below, so every number here can be checked independently.

## TL;DR

FLUX.2 Klein scores 23.9 on our photo-anchored scale, where a real photograph is 100; it is the lowest rung of the ladder. In blind pairwise tests on 48 commercial scenes it fell clearly behind Qwen Image 2.0, GPT Image 2 and the reference photographs, and it was the model furthest from the photographs. Use it for drafts and exploration rather than final photoreal images.

## Where it stands

| Rank | Model | Score | 95% interval |
|---|---|---|---|
| #1 | GPT Image 2 | 104 | — |
| — | Real photograph | 100 | — |
| #2 | Qwen Image 2.0 Pro | 80 | 68.6–93.8 |
| #3 | Qwen Image 2.0 | 65 | — |
| #4 | Seedream 4.0 | 57 | 44–70 |
| #5 | FLUX.2 Klein (this card) | 24 | — |

Axis: believability. A real photograph is 100. A score above 100 is the scale's technical ceiling and does not mean parity.

### By use case

| Category | Score | Rank |
|---|---|---|
| Interior | 64 | #5 |
| Product and e-commerce | 49 | #5 |
| Social marketing | 47 | #5 |
| Fashion and beauty | 43 | #5 |
| Street and architecture | 34 | #5 |
| Layout and typography | −21 | #5 |
| Food and beverage | −41 | #5 |
| People and lifestyle | −56 | #5 |

*Category positions come from 6 scenes each — a tendency, not a measurement. A real photograph is 100 on every ruler.*

## When to use as default

Use FLUX.2 Klein for drafts and quick exploration, where the goal is to try many directions before committing. In our blind test it sat clearly below the other models on believability, so treat its images as sketches of an idea.

## When to use as finisher

Not a finisher. Move the chosen direction up the ladder: Qwen Image 2.0 was preferred over it in 68.6% of judgements and GPT Image 2 in 83.3% (vote shares, ties split), with believability gaps of 0.85 and 1.68.

## When to look elsewhere

Avoid it for anything that must pass as a photograph: of all models on the ladder it was furthest from the reference photographs, with a believability gap of 1.59.

## Direct blind tests

| Opponent | Result | Gap (Bradley–Terry, 95% CI) | Preferred in | Judgements |
|---|---|---|---|---|
| Qwen Image 2.0 | Qwen Image 2.0 was preferred over FLUX.2 Klein in 69% of blind comparisons. | −0.85 (interval not published) | 31.4% (ties split, all questions, vote share) | 144 |
| GPT Image 2 | GPT Image 2 was preferred over FLUX.2 Klein in 83% of blind comparisons. | −1.68 (interval not published) | 16.7% (ties split, all questions, vote share) | 144 |
| Real photograph | The real photograph was preferred over FLUX.2 Klein in 74% of blind comparisons. | −1.59 (interval not published) | 26.0% (ties split, all questions, vote share) | 144 |

## Price

| Provider | Price | Checked |
|---|---|---|
| [Black Forest Labs](https://docs.bfl.ml/quick_start/pricing) | from $0.015 / img | 2026-09-25 |
| [fal.ai](https://fal.ai/models/fal-ai/flux-2/klein/9b) | $0.006 / img | 2026-09-25 |
| [DeepInfra](https://deepinfra.com/flux) | $0.015 / img | 2026-09-25 |
| [WaveSpeedAI](https://wavespeed.ai/models/wavespeed-ai/flux-2-klein-9b/text-to-image) | $0.01 / img | 2026-09-25 |

Average: $0.012 / img across 4 providers

Prices are for one 1024×1024 image at standard settings, checked 25 September 2026; the price at the top of the page is the average of the providers listed. Workroom pricing is on the pricing page.

## License

FLUX.2 Klein comes with open weights from Black Forest Labs in two sizes. The larger one, tested here, is released under the FLUX Non-Commercial License; the smaller one is released under Apache 2.0.

## FAQ

### How does FLUX.2 Klein compare with Qwen Image 2.0?

Clearly behind. The believability strength gap is 0.85 in Qwen Image 2.0's favour, and the dataset notes that the three lower rungs separate reliably. Across the three questions Qwen Image 2.0 was preferred in 68.6% of judgements, counted from the published votes with ties split.

### Why was FLUX.2 Klein chosen as the floor of the scale?

The ladder needs a clearly weaker rung to fix the slope of the scale. FLUX.2 Klein was chosen as that floor and in blind tests it separated reliably from Qwen Image 2 and the photographs. Together with them it is one of the three calibration points the scale is fitted on.

### Does a score of 24 mean the images are unusable?

No. The scale measures one thing: how believable an image is as a real photograph, with a photograph fixed at 100. A low score means that, judged blind, the other image was picked as more photographic more often. Drafts, concepts and non-photographic work are not what this scale measures.

### How was FLUX.2 Klein tested?

It was one of four rungs compared blind on 48 commercial scenes in August 2026, paired on every scene with Qwen Image 2, GPT Image 2 and the reference photographs: 288 pairs and 863 judgements in the study, each pair judged three times on prompt adherence, believability and aesthetics.

### Is Workroom affiliated with Black Forest Labs?

No. Workroom is an AI router that gives access to third-party models, including FLUX.2 Klein, and is not affiliated with Black Forest Labs. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0, so every number on this page can be checked against the source dataset.

## Methodology

Between 21 and 25 August 2026 we compared four rungs of our quality ladder blind on 48 commercial scenes: FLUX.2 Klein, Qwen Image 2, GPT Image 2 and the real photographs each scene was written from. Every rung was paired with every other on every scene — 288 pairs and 863 judgements — and each pair was judged three times, without any indication of which image came from which model, on three questions: which image follows the prompt better, which is more believable as a real photograph, and which is more attractive.

- Scenes: 48; generations per scene: 1
- Compared with: Qwen Image 2.0, GPT Image 2, Real photograph
- Pairs: 288; judgements: 863; each pair judged 3 times
- Collected: 2026-08-21 … 2026-08-25
- Questions: “Which image follows the prompt better?” “Which is more believable as a real photograph?” “Which is more attractive?”
- Scale: 0.020835 strength per point; A real photograph is 100.
- Category positions are fitted from the same judgements, 6 scenes per category, on the same scale.

### Caveats

- Prompt adherence is biased against the photograph: each prompt was written by looking at the photograph, and a model executes text literally, while a photograph always contains incidental detail the description does not mention. Raters count that as a miss — which is why the absolute scale is built on believability alone.
- The share of ties is inflated: during collection the “equal” option was pre-selected, so a criterion the rater never touched was recorded as a genuine tie. The defect was fixed on 28 August 2026, after this data was collected. Tier ordering survives it: dropping every tie only widens the gaps.
- All images were shown at 2048 px on the long edge — native output size for the generative tiers; the photographs were downscaled from 30–45 megapixels. On screen the photographs are not softer than the generated images.

Rater agreement is low by the usual yardstick (Fleiss' kappa 0.16–0.20). Half the pairs compare adjacent tiers, where “equal” is an honest answer. Conclusions hold at the level of the whole corpus, not for a single scene, pair or rater.

## Links

- Full study: https://research.everypixel.com/anchor-ladder/
- Votes (CSV, 106 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/votes.csv
- Results (JSON, 5 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/results.json
- Scenes and prompts (CSV, 30 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/scenes.csv
- README (MD, 2 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/anchor-comparison-v3/README.md
- Images (ZIP, 128 MB, All rights reserved — verification only): https://wr.everypixel.com/ghost-research/geo-benchmarks/datasets/anchor-comparison-v3-images.zip

## Citation

Everypixel research (2026). Anchor comparison of generative image models, v3 [Data set]. Ratings collected 21–25 August 2026.

Data license: CC BY 4.0. Images: All rights reserved; ten Unsplash photographs under the Unsplash License.

---
HTML version: https://workroom.everypixel.com/models/flux-2-klein
Last modified: 2026-08-25