SD 3.5
Triple encoder architecture (CLIP-L + CLIP-G + T5-XXL) with a quantized MMDiT transformer and NaN-safe inference.
- Developer: Stability AI
- License: Stability AI Community License
- HuggingFace: stabilityai/stable-diffusion-3.5-large, stabilityai/stable-diffusion-3.5-large-turbo, stabilityai/stable-diffusion-3.5-medium
Variants
| Model | Steps | Size | Notes |
|---|---|---|---|
sd3.5-large:q8 | 28 | 8.5 GB | 8.1B params |
sd3.5-large:q4 | 28 | 5.0 GB | Smaller footprint |
sd3.5-large-turbo:q8 | 4 | 8.5 GB | Fast 4-step |
sd3.5-medium:q8 | 28 | 2.7 GB | 2.5B params, efficient |
Defaults
- Resolution: 1024x1024
- Guidance: 4.0
- Steps: 28 (4 for turbo)
Recommended Dimensions
| Width | Height | Aspect Ratio |
|---|---|---|
| 1024 | 1024 | 1:1 (native) |
| 1152 | 896 | 9:7 |
| 896 | 1152 | 7:9 |
| 1216 | 832 | 19:13 |
| 832 | 1216 | 13:19 |
| 1344 | 768 | 7:4 |
| 768 | 1344 | 4:7 |
Using non-recommended dimensions will trigger a warning. All values must be multiples of 16.
Example
SD 3.5 Large Q8 — 28 steps, seed 2024:
bash
mold run sd3.5-large:q8 "A steampunk clocktower in a Victorian city at sunset, gears and cogs visible through glass walls, dramatic clouds" --seed 2024
Notes
SD 3.5 uses classifier-free guidance — negative prompts are supported. The quantized MMDiT includes NaN-safe checks for numerical stability.