What happened to SD 2.
Stability released SD 2.0 and then 2.1 in late 2022. They swapped SD 1.5’s CLIP text encoder for the larger OpenCLIP ViT-H and came in two kinds: a 768 model trained with v-prediction, and a 512 base model.
SD 2 never caught on the way SD 1.5 and SDXL did, and most fine-tunes and LoRAs were made for those two. Stability has since deprecated 2.0 and 2.1 and removed their repos, citing newer architectures and upcoming EU AI Act requirements. The community copied the files to sd2-community in November 2025.
Run it for an old workflow, a fine-tune you still like, or out of interest. For new work, SD 1.5 is lighter and SDXL is better.
Files you need.
One checkpoint with the model, text encoder and VAE inside. Pick the kind first: the size you generate at has to match it.
-
768, v-pred5.2 GB Download
v2-1_768-ema-pruned.safetensorsComfyUI/models/checkpoints/ -
512 base5.2 GB Download
v2-1_512-ema-pruned.safetensorsComfyUI/models/checkpoints/
SD 2.0’s 768-v-ema.safetensors is in sd2-community/stable-diffusion-2, along with the inpainting, depth and unCLIP models.
models/
└── checkpoints/
└── v2-1_768-ema-pruned.safetensors
Set it up.
-
Download a checkpoint
From the mirror, into
ComfyUI/models/checkpoints. Press R or restart ComfyUI. -
Load it like SD 1.5
Load Checkpoint, two CLIP Text Encode nodes, Empty Latent Image, KSampler, VAE Decode. The default SD 1.5 workflow works with the checkpoint swapped.
-
Set the size to match
768 × 768 for the 768 models, 512 × 512 for the base models.
-
No yaml needed
ComfyUI tells a v-prediction 768 model from a base model on its own and runs attention in fp32, which SD 2.1 needs. AUTOMATIC1111 wanted a
v2-inference-v.yamlnext to the file; ComfyUI doesn’t.
Settings that work.
- Steps
- 25
- CFG
- 7
- Sampler
- dpmpp_2m
- Scheduler
- karras
- Size
- 768 or 512
- Negative
- yes
- Text encoder
- OpenCLIP ViT-H
- VAE
- from checkpoint
Standard SD settings work. The one that matters is size: the 768 model at 512 pixels, or the base model at 768, gives poor results.
When it goes wrong.
- The old stabilityai download link fails with 401
- Stability removed the repos. Use the sd2-community mirror.
- Blurry or broken images
- Check the size matches the model: 768 × 768 for 768 files, 512 × 512 for base files.
lora key not loaded- An SD 1.5 or SDXL LoRA. SD 2 needs LoRAs made for SD 2.
Why did Stability take SD 2.1 down, and will it come back?
Can someone put Stable Diffusion 2.1 back up? My projects depend on it.
Questions.
Where can I download SD 2.1 now?
From the sd2-community mirror on Hugging Face: stable-diffusion-2-1 for the 768 model and stable-diffusion-2-1-base for the 512 model. The stabilityai repos are gone.
Why was SD 2.1 removed?
Stability deprecated SD 2.0 and 2.1, pointing to newer architectures and compliance requirements under the EU AI Act in 2026.
Does ComfyUI need a yaml file for SD 2.1?
No. ComfyUI detects v-prediction for the 768 models from the weights. The yaml file was an AUTOMATIC1111 requirement.
768 or 512?
Use the 768 model at 768 × 768 and the base model at 512 × 512. Each works poorly at the other size.
Should I use SD 2.1 today?
Only for an existing workflow or fine-tune. SD 1.5 is lighter with more add-ons, and SDXL makes better images on 8 GB.
Sources: sd2-community SD 2.1, SD 2.1 base, Stability’s deprecation note, Hugging Face forum thread.
