sdnext

Generates images from text prompts using HuggingFace Diffusers pipeline architecture with pluggable backend support (PyTorch, ONNX, TensorRT, OpenVINO). The system abstracts hardware-specific inference through a unified processing interface (modules/processing_diffusers.py) that handles model loading, VAE encoding/decoding, noise scheduling, and sampler selection. Supports dynamic model switching and memory-efficient inference through attention optimization and offloading strategies.

Transforms existing images by encoding them into latent space, applying diffusion with optional structural constraints (ControlNet, depth maps, edge detection), and decoding back to pixel space. The system supports variable denoising strength to control how much the original image influences the output, and implements masking-based inpainting to selectively regenerate regions. Architecture uses VAE encoder/decoder pipeline with configurable noise schedules and optional ControlNet conditioning.

Exposes image generation capabilities through a REST API built on FastAPI with async request handling and a call queue system for managing concurrent requests. The system implements request serialization (JSON payloads), response formatting (base64-encoded images with metadata), and authentication/rate limiting. Supports long-running operations through polling or WebSocket for progress updates, and implements request cancellation and timeout handling.

Provides a plugin architecture for extending functionality through custom scripts and extensions. The system loads Python scripts from designated directories, exposes them through the UI and API, and implements parameter sweeping through XYZ grid (varying up to 3 parameters across multiple generations). Scripts can hook into the generation pipeline at multiple points (pre-processing, post-processing, model loading) and access shared state through a global context object.

Provides a web-based user interface built on Gradio framework with real-time progress updates, image gallery, and parameter management. The system implements reactive UI components that update as generation progresses, maintains generation history with parameter recall, and supports drag-and-drop image upload. Frontend uses JavaScript for client-side interactions (zoom, pan, parameter copy/paste) and WebSocket for real-time progress streaming.

Implements memory-efficient inference through multiple optimization strategies: attention slicing (splitting attention computation into smaller chunks), memory-efficient attention (using lower-precision intermediate values), token merging (reducing sequence length), and model offloading (moving unused model components to CPU/disk). The system monitors memory usage in real-time and automatically applies optimizations based on available VRAM. Supports mixed-precision inference (fp16, bf16) to reduce memory footprint.

Provides unified inference interface across diverse hardware platforms (NVIDIA CUDA, AMD ROCm, Intel XPU/IPEX, Apple MPS, DirectML) through a backend abstraction layer. The system detects available hardware at startup, selects optimal backend, and implements platform-specific optimizations (CUDA graphs, ROCm kernel fusion, Intel IPEX graph compilation, MPS memory pooling). Supports fallback to CPU inference if GPU unavailable, and enables mixed-device execution (e.g., model on GPU, VAE on CPU).

Reduces model size and inference latency through quantization (int8, int4, nf4) and compilation (TensorRT, ONNX, OpenVINO). The system implements post-training quantization without retraining, supports both weight quantization (reducing model size) and activation quantization (reducing memory during inference), and integrates compiled models into the generation pipeline. Provides quality/performance tradeoff through configurable quantization levels.

Applies spatial conditioning to image generation using auxiliary models (ControlNet) that encode structural information (pose, depth, edges, semantic maps) as additional guidance signals. The system loads ControlNet weights, processes input images through condition extractors (e.g., OpenPose for pose, MiDaS for depth), and injects conditioning into the diffusion process via cross-attention mechanisms. Supports weighted multi-ControlNet stacking for combined constraints.

Loads and applies low-rank adaptation (LoRA) weights and textual inversion embeddings to modify model behavior without full fine-tuning. The system maintains a registry of adapter weights, merges them into the base model's attention layers using low-rank decomposition, and injects custom token embeddings into the text encoder. Supports weighted composition of multiple LoRAs and dynamic enable/disable without model reloading.

Provides pluggable sampler implementations (DDPM, DDIM, Euler, DPM++, Heun, etc.) with configurable noise schedules (linear, quadratic, karras, exponential) that control the denoising trajectory. The system abstracts sampler selection through a registry (modules/sd_samplers_diffusers.py), allowing users to trade off between speed (fewer steps) and quality (more steps) with different convergence characteristics. Each sampler implements different noise prediction strategies and step scaling algorithms.

Automatically discovers Stable Diffusion model checkpoints in configured directories, extracts metadata (architecture, training data, VAE, clip version), and maintains an in-memory registry for fast switching. The system uses file hashing and metadata caching to avoid re-parsing large checkpoint files, supports multiple checkpoint formats (.ckpt, .safetensors, .pt), and integrates with HuggingFace model hub for automatic downloads. Implements lazy loading to defer model instantiation until first use.

Encodes images to latent space and decodes latents back to pixel space using Variational Autoencoder models. The system supports multiple VAE implementations (standard, VAE-FT, VAE-MSE), configurable precision (fp32, fp16, bf16), and optimization strategies (attention slicing, memory-efficient attention, tiling for large images). VAE selection is decoupled from base model, allowing custom VAE substitution for quality tuning.

Processes text prompts through CLIP text encoder to generate embeddings used as conditioning signals for image generation. The system handles tokenization (splitting prompts into tokens), manages token limits (typically 77 tokens for CLIP), supports weighted prompt syntax (e.g., '(concept:1.5)' for emphasis), and integrates custom token embeddings (textual inversion). Implements prompt weighting through cross-attention scaling and token-level guidance.

Enlarges generated or input images using configurable upscaling algorithms (Real-ESRGAN, SwinIR, BSRGAN, Lanczos, etc.). The system maintains a registry of upscaler models, applies them sequentially or in parallel, and supports chaining multiple upscalers. Implements tiling-based upscaling for memory efficiency on large images and integrates upscaling as a post-processing step in the generation pipeline.

Generates video sequences using specialized pipelines (AnimateDiff, Deforum, frame-by-frame diffusion) that maintain temporal consistency across frames. The system supports motion control through optical flow guidance, implements frame interpolation for smooth playback, and allows keyframe-based animation where specific frames are generated and intermediate frames are interpolated. Integrates with image generation pipeline for consistent styling across video.