Ana içeriğe geç

Stable Diffusion 3 and 3.5

Model Source Revision Weight license Parity (CPU, fp32)
sd3-medium stabilityai/stable-diffusion-3-medium-diffusers ea42f8c Stability AI Non-Commercial Research Community License Verified
sd3.5-medium stabilityai/stable-diffusion-3.5-medium b940f67 Stability AI Community License Verified
sd3.5-large stabilityai/stable-diffusion-3.5-large ceddf0a Stability AI Community License Verified
sd3.5-large-turbo stabilityai/stable-diffusion-3.5-large-turbo ec07796 Stability AI Community License Verified

Catalog: stable-diffusion-3 · Models: sd3-medium, sd3.5-medium, sd3.5-large, sd3.5-large-turbo · Task: t2i

Architecture. Three frozen text encoders—CLIP-L, CLIP-G and T5-XXL—produce a joint sequence of 77 CLIP tokens and 256 T5 tokens plus a 2048-wide pooled CLIP embedding. The SD 3.0 medium checkpoint uses a 24-block MMDiT without q/k normalization. The SD 3.5 medium checkpoint uses MMDiT-X, with dual attention in its first 13 blocks; the large and large-turbo checkpoints use the same 38-block MMDiT architecture. Large-turbo is distilled for 4 steps without guidance. A 16-channel VAE with a shift factor decodes the latents. The flow-match Euler scheduler uses shift 3.

Usage. All four repositories are gated. Accept the repository terms and provide HF_TOKEN when downloading:

HF_TOKEN=... python -m imagehub.pretrained sd3-medium checkpoints/sd3-medium
from imagehub import pipeline

with pipeline("t2i", backend="sd3-medium", model="checkpoints/sd3-medium", device="mps") as pipe:
    result = pipe("a cat, oil painting", seed=42, negative_prompt="blurry")
    result.images[0].save("cat.png")

Settings. SD 3.0 medium defaults to 28 steps and guidance 7.0; SD 3.5 medium defaults to 40 steps and guidance 4.5; large defaults to 28 steps and guidance 3.5; large-turbo defaults to 4 steps and guidance 0.0. All four default to 1024×1024, and width/height must be multiples of 16. The other supported options are negative_prompt, num_images, seed, latents and max_sequence_length (256). Unrecognized settings raise an error.

Seed compatibility. For the validated Alpha and Sigma checkpoints, CPU runs with the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). PixArt-LCM has exact tiny-snapshot coverage; its real-checkpoint run is pending. MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. There is no safety checker or invisible watermark. Img2img, other schedulers and fp16/bf16 parity are not in this release.

License. The weights use the Stability AI Community License; see the license. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.