How to use from the
Use from the
Diffusers library
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Piotr1686/sd21-base-acutance", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

Stable Diffusion 2.1-base β€” byte-exact mirror

This repository is a byte-exact mirror of the stable-diffusion-2-1-base checkpoint. It exists because the original upstream is gone, and it is published so that a downstream application can keep verifying what it downloads.

Nothing here is modified. No fine-tuning, no re-quantisation, no re-serialisation. Every file is reproduced bit for bit; the SHA256 table below is the proof.

Why this mirror exists

As of 2026-09-03, an authenticated query to the Hugging Face API returns missing (not gated) for stabilityai/stable-diffusion-2-1-base and for stabilityai/stable-diffusion-2, and the Stability-AI/stablediffusion repository is gone from GitHub as well. The whole upstream was withdrawn β€” weights, model card and the license document alike.

Note for anyone diagnosing a similar problem: HTTP 401 from the Hub does not distinguish "gated" from "does not exist". Only an authenticated query settles it.

Provenance of the bytes

Copied from Manojb/stable-diffusion-2-1-base
Revision 0094d483a120f3f33dafbd187ea4aa60d10de75c
Checked 2026-09-03
Files 11 (plus model_index.json and feature_extractor/, kept for pipeline compatibility)

The consuming application pins hashes, not the repository id β€” the address is only a source of bytes, and the verdict is issued by SHA256. That is the reason a mirror is a safe substitute at all: the same bytes at a different address are still the same artifact.

Independent corroboration

A hash table proves that this copy matches the repository it was copied from. On its own it says nothing about whether that repository matches what Stability AI published β€” the upstream is gone, so there is no reference copy left to check against, and the claim "byte-exact" would otherwise rest on trusting a single intermediary.

Two mirrors settle it. Both were created before the upstream was withdrawn and before the repository these bytes were copied from even existed, so neither can be a copy of it:

mirror repo created revision files matching
glowforge-dev/stable-diffusion-2-1-base-custom 2023-01-29 c4216a4232677eeac33ef9049692aa9e0fcce578 11/11
LanguageMachines/stable-diffusion-2-1-base 2023-06-28 2579f99c90b5a3019fc115ee9566d5ec4f17fccf 11/11

Checked 2026-09-05 against the Hub API: lfs.sha256 for the three weight files (that field is the SHA256 of the content) and locally computed git blob ids for the eight small ones. No second copy was downloaded β€” the check costs nothing to repeat.

Five further mirrors also match, and are deliberately not listed: they were created after the withdrawal, so they may descend from the same intermediary. They are consistent, not independent β€” and a witness count is only worth what its weakest member is.

SHA256 of every file the application loads

file sha256 bytes
scheduler/scheduler_config.json 11ac5627d7df0fa344b875c4b5722b1767a8a2aa1684c2cf8b4d614300127234 346
text_encoder/config.json 7e33bfc475f6cf76e69b28b101993ca834d1f22ff4d35caefde32d63d4f42fba 613
text_encoder/model.safetensors cce6febb0b6d876ee5eb24af35e27e764eb4f9b1d0b7c026c8c3333d4cfc916c 1 361 597 018
tokenizer/merges.txt 9fd691f7c8039210e0fced15865466c65820d09b63988b0174bfe25de299051a 524 619
tokenizer/special_tokens_map.json f118ab3a983206e4f32583448de6bd6aae4ee21869135cef1f5848a753cdaab6 460
tokenizer/tokenizer_config.json d562b18a0c6d32a168fce3d7a342fce64579ebacffb98fb71b0affc494fc32d8 807
tokenizer/vocab.json e089ad92ba36837a0d31433e555c8f45fe601ab5c221d4f607ded32d9f7a4349 1 059 962
unet/config.json ce0c6d379e3b1d3e1f79338de70c80a1e36f17a6372439a8249b6fb0dfa1b608 911
unet/diffusion_pytorch_model.safetensors 6dfae3e5f7d459b50f4b0850ead945972c75bb0e1897628933e169eb43974214 3 463 726 498
vae/config.json 424117cb534ce03497c41305ed868980123917b2b6abba4bbaa615e968772903 553
vae/diffusion_pytorch_model.safetensors a1d993488569e928462932c8c38a0760b874d166399b14414135bd9c42df5815 334 643 276

Verify a local copy with sha256sum, or with any tool of your choice.

License β€” read this part

The weights are licensed under the CreativeML Open RAIL++-M License, and LICENSE in this repository carries its full text, including the use restrictions in Attachment A. Those restrictions travel with the model: if you redistribute it or a derivative, you must pass them on.

⚠ The bundled document is not the exact file that once shipped with SD 2.1, and saying so plainly matters more than looking tidy. The SD 2.1-base model card pointed at stabilityai/stable-diffusion-2/LICENSE-MODEL (Open RAIL++-M, dated 2022-11-24), and that file is no longer published by anyone β€” of eight SD 2.1 mirrors checked on 2026-09-03, none carried a license text. What you get here instead is the same license from the same licensor in its published form, dated 2023-07-26; its body is generic and makes no mention of any particular model. Substituting it visibly is honest; substituting it silently would not be.

Document CreativeML Open RAIL++-M, 2023-07-26
SHA256 19b6998b569b53ac1fc2158a8a3202c8699a9a4605b47075715d9c96be7fb6d0
Size 14 109 bytes

Attribution

Original model by Stability AI. This repository claims no authorship over the weights β€” it only keeps them reachable and verifiable.

Not affiliated with, endorsed by, or sponsored by Stability AI. The name of the model is used here only to identify which checkpoint these files are, as the license requires the attribution to be kept and forbids any suggestion of endorsement.

Downloads last month
16
Safetensors
Model size
0.9B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support