A familiar WebUI workflow
You work with prompts, checkpoints, samplers, seeds, image dimensions, txt2img, and img2img in the form-and-tab interface inherited from A1111.
Project README ↗Stable Diffusion WebUI Forge gives you a familiar browser interface for local image generation, with flexible model loading and memory controls for demanding workflows.
Install Forge, add your models, generate your first image, and grow into advanced workflows at your own pace.
Forge is a WebUI platform. It is not a Stable Diffusion model, a hosted image service, or a renamed copy of ComfyUI.
You work with prompts, checkpoints, samplers, seeds, image dimensions, txt2img, and img2img in the form-and-tab interface inherited from A1111.
Project README ↗Forge adds its own controls for model storage precision, GPU weight allocation, swap method, and swap location. These settings change how a workload fits; they do not guarantee a speed increase on every machine.
Explore the interface code ↗Forge uses selected ideas and attributed code from ComfyUI, but the maintainer explicitly states that Forge does not use ComfyUI as its backend. The two products also present very different workflows.
Read the architecture notes ↗Forge runs the WebUI, but you still need compatible model files. The model family, precision, and supporting files determine what the interface can load and how much memory the job needs.
See what first run needs →The exact model and feature compatibility depends on the Forge commit, model files, and extensions you run. These are the core jobs the interface is built around.
Choose a checkpoint, enter a prompt, set dimensions and sampling controls, then generate directly from the txt2img tab.
Use img2img and inpainting to change an image, repair a selected region, or guide how far the result moves from the source.
Forge includes ControlNet integration and related preprocessors in the current source. Support still varies by model family and revision.
Load standard or low-bit model formats and decide how much GPU memory is reserved for weights versus inference work.
A “low VRAM” label is not enough to predict whether a job will work. The model, precision, resolution, batch size, extensions, and memory settings all change the result.
Model weights take part of the available GPU or system memory.
Generation needs working space beyond the memory occupied by weights.
Moving data to CPU or shared memory can make a job fit, but it can also make it slower.
CONCEPTUAL ONLY · ACTUAL BEHAVIOR VARIES BY MODEL, PRECISION, GPU, CONFIGURATION, AND COMMIT.
Forge has a lot of controls, but you do not need to learn them all at once. Choose the route that matches your current state.
Choose between the documented Windows archive and the advanced Git route before downloading anything.
Choose an install path02ALREADY INSTALLEDUnderstand the console, localhost, model files, and the clean baseline to establish before adding extensions.
Open the first-run guide03CHOOSING A UIDecide between a form-and-tab WebUI, a node graph, a canvas-first app, or a multi-UI manager.
Compare local image UIs04RETURNING USERLook up launch arguments, file locations, interface terms, offline behavior, API routes, and licensing.
Browse the referenceThese are the details that prevent most wrong downloads, mismatched expectations, and avoidable setup work.
Use the installation guide or open the lllyasviel/stable-diffusion-webui-forge repository. Download buttons on this site send you to the project’s GitHub pages.
No. Forge is the interface and runtime. It includes sd, xl, and flux presets, while checkpoints, UNets, VAEs, text encoders, and LoRAs are added separately when a model family requires them. Use the Forge UI glossary when those labels are unfamiliar, then start with the first-run guide.
No. In the default setup, a local process starts the interface and your browser opens its localhost address. First run can still download dependencies, and model downloads come from their own providers. Do not expose the local server to a network unless you understand the security implications.
It may improve a particular workload, but there is no honest universal percentage. Compare the same model, seed, sampler, steps, dimensions, precision, extensions, commit, and warm-up state. The maintainer’s performance-reporting guidance requires this context for a useful comparison.
The easiest route is the Windows package for NVIDIA CUDA. Dedicated guides also cover Linux, AMD with DirectML or ROCm, and macOS. Start with the installation page and choose the route that matches your system.
Yes. Forge is published under GNU AGPL Version 3. Use the license guide to understand local use, modified network hosting, redistribution, generated output, models, and extensions.
This guide installs the original lllyasviel repository. Forge Neo, reForge, and Forge Classic are separate downstream projects with their own owners, files, features, and support state. Do not use instructions for one as if they applied to another.
The original README documents a Windows one-click package. The handoff is short, but each step has a distinct purpose.
You can review the release notes and choose the package that matches your CUDA and PyTorch setup.