Curiosity · AI Model
Stable Diffusion 3.5 Large
Stable Diffusion 3.5 Large is Stability AI's 8-billion-parameter Multimodal Diffusion Transformer (MMDiT). It was released with open weights and a community licence, giving artists and engineers a self-hostable alternative to closed APIs. It improves prompt adherence, typography, and composition versus SDXL, and is compatible with the vast LoRA / ControlNet ecosystem.
Model specs
- Vendor
- Stability AI
- Family
- Stable Diffusion 3
- Released
- 2024-10
- Context window
- 1 tokens
- Modalities
- text, vision
Strengths
- Open weights under the Stability Community License
- Strong prompt adherence and typography
- Extensive LoRA / ControlNet ecosystem
- Runs on 16–24 GB consumer GPUs with quantisation
Limitations
- No token context — sampling cost scales with steps × resolution, not tokens
- Community licence restricts very-large-scale commercial use — read terms
- Photorealism lags FLUX 1 pro and Midjourney v6.1 for portraits
- Requires self-hosted GPU infrastructure for best results
Use cases
- Self-hosted creative pipelines in ComfyUI / Automatic1111
- LoRA-driven brand and character workflows
- ControlNet composition and pose-guided generation
- Offline production art on prosumer hardware
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| GenEval composition | ≈0.71 | 2024-10 |
| PickScore (human pref) | ≈60% | 2024-10 |
Frequently asked questions
What is Stable Diffusion 3.5 Large?
Stable Diffusion 3.5 Large is Stability AI's 8-billion-parameter Multimodal Diffusion Transformer (MMDiT) text-to-image model. It was released with open weights under the Stability Community License in October 2024.
Can I use it commercially?
Yes, within the Stability Community License — individuals and businesses under a specified annual revenue threshold may use the weights commercially. Larger enterprises need a commercial agreement with Stability AI.
What hardware do I need?
SD 3.5 Large comfortably runs on a 24 GB GPU in FP16, or a 16 GB GPU with 8-bit quantisation. Lower-tier variants (SD 3.5 Medium and Turbo) run on even smaller GPUs.
How does it compare to FLUX 1?
FLUX 1 [pro] generally leads on photorealism and composition, while SD 3.5 benefits from a much larger LoRA / ControlNet community and a permissive self-host story. Pick SD 3.5 when you want full control; pick FLUX when you want the highest base quality.
Sources
- Stability AI — SD 3.5 announcement — accessed 2026-04-20
- Hugging Face — stabilityai/stable-diffusion-3.5-large — accessed 2026-04-20