Curiosity · AI Model
Black Forest Labs FLUX.1 [dev]
FLUX.1 [dev] is the open-weight sibling of FLUX.1 [pro] — a 12-billion-parameter diffusion transformer released under a non-commercial research licence. It delivers near-[pro] quality on a single prosumer GPU and anchors the community LoRA ecosystem around FLUX.
Model specs
- Vendor
- Black Forest Labs
- Family
- FLUX.1
- Released
- 2024-08
- Context window
- 1 tokens
- Modalities
- text, vision
Strengths
- Near-[pro] quality with open weights
- Runs on 24 GB GPUs (FP16) or 12 GB with quantisation
- Active community — LoRAs, ControlNets, and IP-Adapters
- Transparent architecture for research
Limitations
- Non-commercial research licence — cannot sell outputs without a separate deal
- Slightly behind [pro] on photorealism
- No token context — cost is compute, not tokens
- Requires self-hosted GPU infrastructure
Use cases
- Self-hosted research and experimentation
- Community LoRA training for characters and styles
- Offline creative pipelines in ComfyUI
- Fine-tuning and distillation experiments
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| GenEval composition | ≈0.74 | 2024-08 |
| Human preference vs SDXL | ≈70% win-rate | 2024-08 |
Frequently asked questions
What is FLUX.1 [dev]?
FLUX.1 [dev] is the open-weight, research-licensed version of Black Forest Labs' 12-billion-parameter FLUX text-to-image diffusion transformer. It is the standard base model for the FLUX community LoRA ecosystem.
Can I use FLUX.1 [dev] commercially?
Not under the default non-commercial licence. For commercial use, either purchase a licence from Black Forest Labs or use FLUX.1 [pro] / [schnell] (schnell is Apache-licensed) depending on your quality and cost needs.
What hardware do I need for FLUX.1 [dev]?
A single 24 GB GPU runs FP16 comfortably. With 8-bit or GGUF quantisation the model fits on 12–16 GB consumer GPUs, albeit with some quality or speed trade-off.
How does it compare with Stable Diffusion 3.5?
FLUX.1 [dev] generally leads on photorealism and prompt adherence, while SD 3.5 benefits from a larger LoRA and ControlNet ecosystem plus a commercial community licence. Pick based on whether quality or ecosystem breadth matters more.
Sources
- Black Forest Labs — FLUX.1 announcement — accessed 2026-04-20
- Hugging Face — black-forest-labs/FLUX.1-dev — accessed 2026-04-20