whatmodel.sh / image / FLUX.1 [dev]
FLUX.1 [dev]
↗ share
+ contribute
FLUX.1 [dev] is an open-weights image model from Black Forest Labs, released August 2024. Its published size is 12B parameters. Licence: flux-dev-nc.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of FLUX.1 [dev], out of 197 with a score, then more from Black Forest Labs.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 125 |  DALL·E 3 | 44±1 | comparison only | ||
| 126 | FLUX.1 Kontext [dev] | 44±1 | api $0.025/image +local | ||
| 127 | FLUX.1 [dev] | 44±1 | api $0.025/image +local | ||
| 128 |  UniPic2-SD3.5M-Kontext-2B | 44±6 | comparison only | ||
| 129 |  UNO | 43±4 | local | ||
| more from Black Forest Labs | |||||
| 32 | FLUX.2 [max] | 64±1 | api $0.070/image | ||
| 36 | FLUX.2 [pro] | 63±2 | api $0.030/image | ||
| 37 | FLUX.2 [flex] | 62±2 | api $0.050/image | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #80 of 126 · 90% of its score · 9 tests
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), FLUX.1 Kontext [dev] here, though they rank higher overall. Trails UNO here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena text-to-image | 969.4 ✎ | ▂▃▃▄▇█▄▂▄▂·▂ beats 4/75 | 28% |
| T2I-ReasonBench (overall) | 54.9% ✎ | ▅▂█▂▂▂····▂▂ beats 6/13 | 16% |
| QC-T2I-Bench (English, HCQ total) | 76.2% ✎ | ▄····▄█▄▄█▄█ beats 0/10 | 14% |
| TIIF-Bench v3 (short prompts, TIIF-Evaluator) | 72.6% ✎ | ▂█····▄▄▄··▂ beats 4/11 | 13% |
| MMMG knowledge-image generation (MMMG-Score, average over education levels) | 19.1% ✎ | ▃··▃·█▃▃▅··▃ beats 7/9 | 5.4% |
| GenExam (relaxed overall, GPT-5 judge) | 17.6% ✎ | ▆▆·▂▂····▂▂█ beats 1/13 | 4.8% |
| UltraText Bench (Composite, bilingual dense text rendering) | 18.2% ✎ | ▁▃▂█▃▁▂▃·▁·▁ beats 1/22 | 4.3% |
| WISE (WiScore overall, GPT-4o judge) | 50.0% ✎ | ▄·▆▂▆█▄▆▂▄·▄ beats 14/22 | 3.9% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| HPSv3 image generation benchmark (Overall HPSv3 score) | 10.4300 ✎ | ▃·····▃··▃·█ beats 4/5 | 0.6% |
other tests2 not in a task · 8.9% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VVRBench accuracy (10,000 verifiable synthetic-scene tasks) | 3.9% ✎ | ▃·█▅▃·▅▃···▃ beats 4/10 | 8.1% |
| Knowledge Visualization T2I (Brief Caption, English overall) | 33.9% ✎ | ▅·▃·▃··▃·█▃▃ beats 5/9 | 0.8% |
$ where to run it
Through an API
api · Replicate · $0.025/image · replicate.com ↗ · price sourceUSD per output image. Text-to-image; one output image, standard FLUX.1 dev endpoint. Replicate output/character billing; hardware-time rates are not added to this tariff. Taxes and separate services excluded. Bertrande, https://bertrande.com/en/ai/replicate/black-forest-labs--flux-dev; observed 2026-09-20T01:44:15Z. CC BY 4.0; units normalized and model identifier mapped.
Other ways to run it
version
- Run it as published · needs ~7.7 GB, NVIDIA only · from huggingface.co/blog
uv run --with diffusers --with transformers --with accelerate --with torch --with sentencepiece --with bitsandbytes python -c "import torch; from diffusers import FluxPipeline, PipelineQuantizationConfig; q=PipelineQuantizationConfig(quant_backend='bitsandbytes_4bit', quant_kwargs={'load_in_4bit':True,'bnb_4bit_quant_type':'nf4','bnb_4bit_compute_dtype':torch.bfloat16}, components_to_quantize=['transformer','text_encoder_2']); p=FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-dev', quantization_config=q, torch_dtype=torch.bfloat16).to('cuda'); p('a cat holding a sign', num_inference_steps=50, guidance_scale=3.5).images[0].save('out.png')"
Official weights and documentation: huggingface.co/black-forest-labs ↗
$ how the score was worked out · LMArena leaderboard dataset 28%, T2I-ReasonBench: Benchmarking… 16%, QC-T2I-Bench: Scalable Questi… 14% and 8 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ published speed measurements · 2
Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.
| NVIDIA L20 (48GB; single GPU) | bf16 (FluxPipeline dtype=torch.bfloat16, no quantization) | 26.36 s/image | diffusers FluxPipeline (ParaAttention guide baseline; diffusers/torch versions not stated) · black-forest-labs/FLUX.1-dev; 1024x1024; 28 inference steps; prompt 'A cat holding a sign that says hello world'; 1 GPU; no optimizations (line 16-18); wall time from pipe() call Baseline row of the Benchmarks table; the prose (line 18) states 1024x1024, 28 steps, 26.36 s. Contributed to the diffusers docs by the ParaAttention author (PR #10544, Jan 2025); datacenter GPU, not a consumer card. | source ↗ |
| NVIDIA L20 (48GB; single GPU) | FP8 dynamic quantization (torchao float8_dynamic_activation_float8_weight on transformer; float8_weight_only on text encoder per the fp8 section) | 13.4 s/image | diffusers FluxPipeline + torchao FP8 dynamic quantization (ParaAttention guide; versions not stated) · black-forest-labs/FLUX.1-dev; 1024x1024; 28 inference steps; 1 GPU; table row 'FP8 DQ' (no First Block Cache, no context parallelism); whether torch.compile was applied for this specific row is not stated Table row 'FP8 DQ'. The fp8 code example additionally shows FBCache and torch.compile; only the combined FBCache+FP8 result (7.56 s) is described in prose, so compile use for this row is unknown. Contributed to the diffusers docs by the ParaAttention author (PR #10544, Jan 2025); datacenter GPU, not a consumer card. | source ↗ |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json