Highlights for 2024-10-23

A month later and with nearly 300 commits, here is the latest SD.Next update!

Workflow highlights

• Reprocess: New workflow options that allow you to generate at lower quality and then
  reprocess at higher quality for select images only or generate without hires/refine and then reprocess with hires/refine
  and you can pick any previous latent from auto-captured history!
• Detailer Fully built-in detailer workflow with support for all standard models
• Built-in model analyzer
  See all details of your currently loaded model, including components, parameter count, layer count, etc.
• Extract LoRA: load any LoRA(s) and play with generate as usual
  and once you like the results simply extract combined LoRA for future use!

New models

• New fine-tuned CLiP-ViT-L 1st stage text-encoders used by most models (SD15/SDXL/SD3/Flux/etc.) brings additional details to your images
• New models:
  Stable Diffusion 3.5 Large
  OmniGen
  CogView 3 Plus
  Meissonic
• Additional integration:
  Ctrl+X which allows for control of structure and appearance without the need for extra models,
  APG: Adaptive Projected Guidance for optimal guidance control,
  LinFusion for on-the-fly distillation of any sd15/sdxl model

What else?

• Tons of work on dynamic quantization that can be applied on-the-fly during model load to any model type (you do not need to use pre-quantized models)
  Supported quantization engines include BitsAndBytes, TorchAO, Optimum.quanto, NNCF compression, and more…
• Auto-detection of best available device/dtype settings for your platform and GPU reduces neeed for manual configuration
  Note: This is a breaking change to default settings and its recommended to check your preferred settings after upgrade
• Full rewrite of sampler options, not far more streamlined with tons of new options to tweak scheduler behavior
• Improved LoRA detection and handling for all supported models
• Several of Flux.1 optimizations and new quantization types

Oh, and we’ve compiled a full table with list of top-30 (how many have you tried?) popular text-to-image generative models,
their respective parameters and architecture overview: Models Overview

And there are also other goodies like multiple XYZ grid improvements, additional Flux ControlNets, additional Interrogate models, better LoRA tags support, and more…
README | CHANGELOG | WiKi | Discord