V2.1.4beta1

What's Changed

• New minimalistic UI and added testing by @Aatricks in #20
• Feat/optimizations stability by @Aatricks in #21

Full Changelog: V2.1.3...V2.1.4beta1

V2.1.3

What's changed

• NVFP4 (4-bit) Weight-Only Quantization Support by @Aatricks
  • Implementation of 4-bit quantization for ~75% reduction in weight memory usage.
  • Integrated support for Flux2 (Transformer + Klein Text Encoder), SDXL, and SD1.5 architectures.
  • Optimized runtime dequantization to FP16/BF16 during the forward pass via comfy_cast_weights.
  • Automated layer selection targeting weights >4096 elements to balance compression and quality.
  • Added weight_quantization configuration to API, Context, and Pipeline for granular memory
    control.

Full Changelog: V2.1.2...V2.1.3

V2.1.2

What's Changed

• Fix: reload base model before HiresFix + stability & UX improvements (Flux2, VAE, ADetailer, settings history) by @Aatricks in #19
• Pipeline Stability & HiresFix Improvements by @Aatricks
  • Fixed critical race conditions in VAE encoding during SDXL+Refiner workflows by enforcing blocking transfers.
  • Implemented logic to reload the base model before HiresFix when a refiner is used, preventing latent corruption.
  • Ensured ADetailer explicitly uses the base model instead of the refiner for text-guided crop enhancements.
  • Reverted non-essential SDXL changes to resolve regressions in Attention and conditioning modules.
• Settings Persistence & History Management by @Aatricks
  • Implemented backend storage for settings history and last-used seeds.
  • Added a collapsible "Settings History" section to the UI for quick restoration of previous configurations.
  • Integrated image import functionality directly into the GenerationSettings component.
• Flux.2 & Core Model Optimization by @Aatricks
  • Enhanced torch.compile integration with support for callables and improved logging.
  • Fixed RoPE feature dimension alignment and added padding adjustments for Flux2.
  • Improved FP8 quantization fallback logic for models lacking diffusion submodules.
  • Added validation for model file existence (safetensors/pt) in the downloader.
• Image Processing & Batch Limits by @Aatricks
  • Introduced LD_MAX_IMAGES_PER_GROUP to control processing limits and implemented chunking for large pipeline requests.
  • Updated AutoHDR to properly handle RGBA images (preserving alpha) and added fallbacks for missing LCMS or failed ICC transforms.
  • Added telemetry for batch limit configuration.
• Testing & Infrastructure by @Aatricks
  • Restructured the test suite into clear e2e, integration, and unit categories.
  • Fixed frontend runtime crashes by importing missing Button and ImageMetadata types.

Full Changelog: V2.1.1...V2.1.2

V2.1.1

What's changed

• Enhanced Preview and Message Management by @Aatricks
  • Added generation ID handling to improve preview message tracking and management.
  • Implemented configurable preview fidelity settings (format and quality) in AppInstance.
• Model Compilation and Image Processing Optimizations by @Aatricks
  • Updated compile_model to default to 'max-autotune-no-cudagraphs' for better performance and stability.
  • Introduced in-memory image byte storage in ImageSaver to reduce disk I/O during API responses.
  • Added color utility functions for linear to sRGB conversion and Reinhard tonemapping.
• Improved HiresFix and SDXL Support by @Aatricks
  • Enhanced HiresFix with support for size conditioning specific to SDXL models.
  • Refined handling of denoise and CFG parameters during the upscaling process.
• Comprehensive Img2Img Enhancements by @Aatricks
  • Expanded API to support image uploads via local file paths, data URLs, and raw Base64 strings.
  • Implemented robust image saving with automatic format conversion and size limit enforcement.
  • Added a request filename prefix feature for improved output file organization.
• Pipeline Robustness and Bug Fixes by @Aatricks
  • Enhanced tensor handling in sampling utilities and the multiscale manager for better stability.
  • Fixed refiner prompt usage for per-sample HiresFix/Adetailer when using SDXL or Flux models.
  • Added missing flux2 tokenizer merges configuration.
  • Improved error handling for non-tensor outputs in VAE and Pipeline modules.
• Tooling and Distribution Updates by @Aatricks
  • Added a dedicated downloader for Flux models to streamline setup.
  • Included the frontend dist folder in the repository.
• Expanded Integration and Unit Testing by @Aatricks
  • Added tests for FP8 quantization and torch.compile compatibility.
  • Introduced comprehensive integration tests for batched processing and high-payload img2img requests.

Full Changelog: 2.1.0...V2.1.1

2.1.0

What's Changed

• Server first by @Aatricks in #18
• Added Img2Img support by @Aatricks
• Added canny Controlnet support by @Aatricks
• Replaced Streamlit and Gradio ui with a new react one by @Aatricks

Full Changelog: V2.0.0...2.1.0

V2.0.0

What's Changed

• Full Flux.2 Klein 4B Distilled Support by @Aatricks in
  #17
  • Implementation of the Flux2 transformer and Qwen-based text encoder.
  • Optimized text conditioning with attention masks, text normalization, and
    vector input support.
  • Resolution-dependent timestep scheduling and model sampling shift
    configurations.
  • Aggressive VRAM management and partial model loading for consumer hardware.
  • Fixed positional embeddings and VAE decoding logic specific to the Klein
    architecture.
• SDPA Backend Priority Management (SpargeAttn > SageAttention > Xformers) by
  @Aatricks
• Enhanced SDXL Condition Processing for improved prompt adherence by @Aatricks
• Optimized Model Reuse Logic to prevent redundant device transfers and reloads by
  @Aatricks
• Streamlined CI Workflow with improved error reporting and expanded
  unit/integration tests by @Aatricks
• Revamped Streamlit UI for synchronized resolution management and Flux2 presets by
  @Aatricks
• Fixed random seed generation to comply with PyTorch limits across all modules by
  @Aatricks
• Improved Img2Img upscale logic and dimension handling for DiT models by @Aatricks

Full Changelog: V1.9.1...V2.0.0

V1.9.1

What's Changed

• Smart cfg by @Aatricks in #11
• ROCm and mps by @Aatricks in #13
• Avoid unnecessary GGUF model unloading after patches by @google-labs-jules[bot] in #16
• Optimize mmap release logic in Quantizer by @google-labs-jules[bot] in #15
• Dynamic Width Scaling in Condition Encoding by @google-labs-jules[bot] in #14
• Various optimizations to reduce device transfers by @Aatricks
• Implemented calculations caching and batching for flux and attention by @Aatricks
• Vectorized tensor indexing for schedulers by @Aatricks
• Implemented dynamic vae tiling based on available VRAM by @Aatricks
• These various improvements lead to about 30% inference speed improvements in SD1.5 scenarios

New Contributors

• @google-labs-jules[bot] made their first contribution in #16

Full Changelog: V1.9.0...V1.9.1

V1.9.0

What's Changed

• Better scheduler (AYS) by @Aatricks in #9
• More models (flux quant levels and SDXL, SD2.5) by @Aatricks in #10
• Implemented deep cache optimization and prompt caching

Full Changelog: V1.8.0...V1.9.0

V1.8.0

What's Changed

• feat: Enable CUDA graphs for faster inference by @Aatricks in #7
• Feature/deepcache by @Aatricks in #8
• Chore : Completely updated the docs

Full Changelog: V1.7.3...V1.8.0

V1.7.3

What's Changed

• Smart batching by @Aatricks in #6
• Implemented requests buffering and generation batching for similar requests in server mode
• Fixed unwanted behaviour on batches in the streamlit UI

Full Changelog: V1.7.2...V1.7.3