Stable Diffusion 1.5

The original 860M UNet latent diffusion model with a single CLIP ViT-L text encoder, trained at 512×512. Small and fast to train, with the largest library of community fine-tunes of any model.

weights: stable-diffusion-v1-5/stable-diffusion-v1-5 ↗
org: Runway / CompVis
modality: image
tasks: text-to-image
license: CreativeML Open RAIL-M
released: 2022-10-20not verified
native output: 512×512not verified
total params: 1.07B
model.arch: sd15

Components

role	model	params	size	dtype	trained
UNet	SD 1.5 UNet (EMA weights) stable-diffusion-v1-5/stable-diffusion-v1-5/unet/diffusion_pytorch_model.safetensors diffusers.UNet2DConditionModel The file Diffusers loads by default. The repo also has fp16 and .bin copies, non-EMA weights (diffusion_pytorch_model.non_ema.safetensors) and single-file checkpoints at the root.	859.5M 859,520,964	3.44 GB	fp32	yes
Text encoder	CLIP ViT-L/14 text stable-diffusion-v1-5/stable-diffusion-v1-5/text_encoder/model.safetensors transformers.CLIPTextModel The count includes 77 int64 position ids stored as a tensor; 123,060,480 weights.	123.1M 123,060,557	492 MB	fp32+int64	no
Tokenizer	CLIP BPE (77 tokens) stable-diffusion-v1-5/stable-diffusion-v1-5/tokenizer transformers.CLIPTokenizer	—	—	—	—
VAE	SD VAE (kl-f8) stable-diffusion-v1-5/stable-diffusion-v1-5/vae/diffusion_pytorch_model.safetensors diffusers.AutoencoderKL No scaling_factor in its config, so the Diffusers default 0.18215 applies.	83.7M 83,653,863	335 MB	fp32	no
Safety checker	CLIP ViT-L/14 image classifier stable-diffusion-v1-5/stable-diffusion-v1-5/safety_checker StableDiffusionSafetyChecker Shipped in the repo but not loaded: ai-toolkit passes safety_checker=None.	—	—	—	no
total		1.07B	4.27 GB

Latent space

spatial

8×

channels

patch

1×1

autoencoder: SD VAE (kl-f8)
pixels per token: 8×8
notes: A UNet has no patch embedding; each latent pixel is one position. The token counts below are latent pixels at full resolution. Attention runs at the 1×, 2× and 4× downsampled levels.

input	latent (c×t×h×w)	tokens
512×512	4×1×64×64	4,096	native
768×768	4×1×96×96	9,216

Architecture

Levels: 4 (320, 640, 1280, 1280 channels), 2 res blocks each
Transformer depth: 1 block at each of the first 3 levels and in the mid block; none at the lowest
Heads: 8 per attention layer (head size 40 / 80 / 160)
Text conditioning: Cross-attention on the final CLIP-L hidden states (768-d), 77 tokens
Objective: Epsilon prediction, scaled-linear betas (0.00085–0.012), 1000 steps

In AI Toolkit

model.arch: sd15
UI label: SD 1.5 (image)
model.name_or_path: stable-diffusion-v1-5/stable-diffusion-v1-5
source: toolkit/stable_diffusion_model.py
toolkit/train_tools.py
toolkit/config_modules.py
toolkit/util/get_model.py
toolkit/sampler.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections

UI defaults

noise_scheduler / sampler: ddpm / ddpm
sample size: 512×512
sample.guidance_scale: 6
quantize / train.timestep_type: hidden in the UI

Specifics

Model code: Handled by the legacy StableDiffusion class, not a model plugin. Loads a StableDiffusionPipeline.
Loading: A Hub id or folder loads with from_pretrained (fp32 safetensors, no fp16 variant) and no safety checker. A local .safetensors or .ckpt file loads with from_single_file. vae_path swaps the VAE.
Resolution: Buckets snap to multiples of 16.
Text encoder: Frozen by default. Uses the final hidden states (no clip skip); long prompts can be chunked in 77-token windows.
Noise schedule: ddpm uses the scaled-linear SD schedule with epsilon prediction. is_v_pred switches to v-prediction for v-pred fine-tunes.
Network: Conv (LoCon) layers are available alongside linear LoRA; LoRA targets UNet2DConditionModel.
Saving: LoRAs use kohya-style keys (lora_unet_, lora_te_) that ComfyUI and A1111 load. Full fine-tunes save a single-file LDM checkpoint by default (save_format diffusers for a folder).
Metadata base version: sd_1.5

Example config

not verified

job: extensionconfig:  name: "my_sd15_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          resolution: [512, 768]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "ddpm"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "stable-diffusion-v1-5/stable-diffusion-v1-5"        arch: "sd1"      sample:        sampler: "ddpm"        sample_every: 250        width: 512        height: 512        guidance_scale: 6        sample_steps: 25        prompts:          - "a bear building a log cabin in the snow covered mountains"