The A1111 fork that natively speaks Flux and publishes its bugs

What it does Forge is a fork of AUTOMATIC1111’s Stable Diffusion WebUI—currently tracking version 1.10.1—that adds a GPU memory management system, faster inference hooks, and a sandbox for experimental features. It natively supports Flux models in quantized formats like GGUF and BitsandBytes NF4, complete with a “GPU Weight” slider and offloading toggles so you can trade generation speed for VRAM without leaving the Gradio interface.

The interesting bit The project wears its Minecraft Forge inspiration openly: it wants to be the modding platform for SD WebUI. The most concrete example is the UnetPatcher system, which lets developers inject UNet modifications—like the FreeU V2 filter shown in the README—in a single self-contained file rather than rewriting the entire inference stack.

Key highlights

• Native Flux support for GGUF Q4_0/Q4_1/Q5_0/Q5_1/Q8_0 and BitsandBytes NF4, with LoRA support for most of those formats.
• GPU weight slider and async swap toggles for offloading layers between GPU, system RAM, and disk.
• A public “Forge Status” table that manually tracks what works, what’s broken (OFT LoRAs, Surface touch pressure), and what’s not yet implemented (Flux ControlNets, Union ControlNets).
• UnetPatcher API for single-file UNet modifications without touching the core diffusion loop.
• Gradio 4-based UI with a Forge Canvas that supports Wacom 128-level pressure.

Caveats

• Several features are explicitly marked broken or pending: OFT LoRAs, Microsoft Surface touch pressure in the canvas, and ControlNet support for Flux/Union models.
• API endpoints for txt2img/img2img are functional but still pending improved Flux support.
• The fastest CUDA 12.4 package may have broken MSVC and non-working xformers.

Verdict Worth a look if you’re hitting VRAM walls with A1111 or want to run quantized Flux models without switching to ComfyUI. Skip it if you rely on OFT LoRAs or need fully stable ControlNet for Flux today.