stable-diffusion-v1-5

Generates photorealistic and artistic images from natural language text prompts using a latent diffusion model architecture. The pipeline encodes text prompts into CLIP embeddings, then iteratively denoises a random latent vector through 50+ diffusion steps guided by the text embedding, finally decoding the latent representation back to pixel space via a VAE decoder. This approach reduces computational cost compared to pixel-space diffusion by operating in a compressed 4x-4x-8x latent space.

Implements classifier-free guidance (CFG) during the diffusion process by computing conditional and unconditional noise predictions, then blending them with a guidance_scale weight to steer generation toward the text prompt. At each denoising step, the model predicts noise for both the text-conditioned and unconditioned (empty prompt) latents, then interpolates: noise_final = noise_uncond + guidance_scale * (noise_cond - noise_uncond). Higher guidance_scale (7.5-15.0) increases prompt adherence at the cost of reduced diversity and potential artifacts.

Enables parameter-efficient fine-tuning via Low-Rank Adaptation (LoRA), where only small rank-decomposed matrices are trained instead of full model weights. LoRA adds trainable weight matrices (A and B) to selected layers, with rank typically 4-64. During inference, LoRA weights are merged into the base model or applied as a separate forward pass. This approach reduces fine-tuning memory from ~24GB (full model) to ~2-4GB (LoRA only) and enables fast adaptation to new styles, objects, or concepts.

Generates new images conditioned on an input image by encoding the image into latents, adding noise according to a strength parameter (0.0-1.0), and then denoising with text guidance. Strength controls how much the output deviates from the input: strength=0.0 returns the input image unchanged, strength=1.0 ignores the input and generates from scratch. Internally, the pipeline skips the first (1 - strength) * num_inference_steps denoising steps, preserving input image structure while allowing variation.

Generates images within masked regions while preserving unmasked areas, enabling targeted image editing. The inpainting pipeline accepts an image, mask (binary or soft), and text prompt. Masked regions are encoded into latents, noise is added, and the diffusion process generates new content in masked areas while keeping unmasked areas fixed. The mask is applied at each denoising step to blend generated and original content. This enables precise control over which image regions are modified.

Processes multiple text prompts in parallel by batching latent tensors and text embeddings through the diffusion loop, with per-sample seed control for reproducibility. The pipeline accepts batch_size > 1, generates unique random latents for each sample (or uses provided seeds), and returns a batch of images in a single forward pass. Seed management uses PyTorch's random number generator state to ensure deterministic output when the same seed is provided.

Accepts a negative_prompt parameter that is encoded into embeddings and used during classifier-free guidance to suppress unwanted visual concepts. The pipeline computes noise predictions conditioned on both the positive prompt and negative prompt, then uses guidance to push the generation away from the negative prompt direction. Internally, negative prompts are concatenated with positive prompts in the batch dimension, requiring 2x text encoding passes (or 1 pass with concatenation) to generate both embeddings.

Encodes text prompts into 768-dimensional CLIP embeddings using a pre-trained CLIP text encoder (trained on 400M image-text pairs). The encoder tokenizes input text (max 77 tokens), passes tokens through a transformer, and extracts the final hidden state as the embedding. These embeddings are then used to condition the diffusion process via cross-attention layers in the UNet. CLIP embeddings capture semantic meaning of text in a space aligned with image features, enabling the diffusion model to generate images matching the text description.

Compresses 512x512 RGB images into 64x64x4 latent tensors using a pre-trained Variational Autoencoder (VAE) encoder, enabling diffusion to operate in a compressed space. The VAE encoder downsamples the image through convolutional blocks with residual connections, producing a latent distribution (mean and log-variance). During generation, the VAE decoder upsamples the denoised latent back to 512x512 RGB pixel space. This compression reduces memory and computation by ~64x compared to pixel-space diffusion.

Conditions the diffusion process on text embeddings via cross-attention layers in the UNet. At each denoising step, the UNet computes self-attention over spatial features and cross-attention between spatial features and text embeddings. The cross-attention mechanism (Q from spatial features, K and V from text embeddings) enables the model to selectively attend to relevant parts of the prompt at each spatial location. This architecture allows fine-grained control over which prompt concepts influence which image regions.

Generates images through 50+ iterative denoising steps, where at each step the model predicts noise added to the latent and subtracts it. The process uses a timestep scheduler (e.g., DDPM, PNDM, Euler) that defines the noise schedule (how much noise to add/remove at each step) and the order of steps. The scheduler controls the trade-off between inference speed (fewer steps, faster but lower quality) and quality (more steps, slower but higher quality). Common schedulers include DDPM (50 steps), PNDM (20 steps), and Euler (20-50 steps).

Loads model weights from safetensors format (a memory-safe serialization format) instead of pickle, preventing arbitrary code execution during model loading. Safetensors uses a simple binary format with explicit type information, enabling safe deserialization without executing Python code. The diffusers library automatically detects and loads safetensors files, falling back to pickle if safetensors is unavailable. This approach reduces security risk when loading untrusted model weights from HuggingFace or other sources.

Supports mixed-precision inference (fp16 or int8) to reduce memory footprint and increase speed, and enables memory-efficient attention implementations (e.g., xFormers, Flash Attention) to reduce attention memory complexity from O(n²) to O(n). Users can enable mixed-precision via `pipe.to('cuda', dtype=torch.float16)` and memory-efficient attention via `enable_attention_slicing()` or `enable_xformers_memory_efficient_attention()`. These optimizations are composable and can be combined for maximum efficiency.