What is Forge

Forge WebUI started out as a modified version of the popular Automatic1111 Stable Diffusion interface. It added some efficiency improvements, some of which came from copying code from ComfyUI as well. In recent versions, it allows you to use a bigger range of models, including the popular new Flux model, letting it follow more complex prompts and create scenes that weren't possible before.

Most of Forge's interface will already be familiar to you if you started with Automatic1111. There is a txt2img (Text-to-Image) tab with options for entering prompts and generating images, an img2img (Image-to-Image) tab for modifying or regenerating images, an Extras tab full of tools and utilities, a PNG Info tab that provides a history by retrieving the prompts and settings from the images you've generated before, and an Extensions tab full of extensions, a valuable collection of plug-ins that add functionality to Forge.

In earlier versions, people weren't sure if it was worth switching to Forge. You could gain some performance improvements compared to Automatic1111, but it would come at a cost of potentially reduced compatibility for some extensions. For a while, it wasn't clear that Forge was going to get continuous development at all. But now, with new versions and a new generation of models available, Forge provides an easy way to start using Flux, while continuing with a familiar interface. Of course, Forge lets you continue using all your older Stable Diffusion models as well.

Forge is faster than Automatic1111 but slower than SwarmUI.

For generations in Flux, I timed a scene (the clown scene shown in a screenshot above) that was 1280x720 resolution, 30 steps, using Flux.1-Dev model. It took one minute and 12 seconds to generate in Forge. The same prompt and settings in SwarmUI took only 46 seconds. I generated the scene on a system with an RTX 3090 graphics card, with 24GB of VRAM, ran it several times in each package, and did not time the first image (the first image generated tends to take about 30 seconds longer because the model is initially being loaded into memory.)

Forge recently came out with a new sampler called "Flux Realistic." The generation time dropped to about one minute and 3 seconds with this sampler. Forge also added a sampler called "Flux Realistic (Slow)" which took 1:48 to generate, which is a full minute longer than Swarm took for the scene, but did produce terrific looking images, with a slight detail enhancement and very good definition on textured surfaces without looking over-sharpened. Seeing these new choices coming out, even as I write this review, is terrific. Even though the tab at the top of the interface is still labelled "Stable Diffusion," Forge WebUI is already a powerful interface for a world that's moving beyond the Stable Diffusion era.

SD.Next has recently announced that it supports Flux also. However, the version I downloaded did not support the full Flux.1-Dev model that I was comparing between Forge and Swarm. Instead, it downloaded a number of smaller diffusers files. It very slowly managed to generate a Flux image from them, but I couldn't do a head-to-head comparison if it doesn't support the same full-size model that the others use.

If you've been using Stable Diffusion for a few years with the Automatic1111 interface, but are looking for a way to upgrade from Stable Diffusion models to Flux, then Forge UI seems like a great place to start. You can keep using a familiar, full-featured interface, but also create more advanced imagery and use more complex and descriptive prompts. Installation instructions are available on Forge's GitHub page.