Run AI Image Generation Without Restrictions

What you need

Forge is a fork of AUTOMATIC1111's WebUI with better memory handling, so it runs on smaller cards than the original. The portable build ships its own Python — you don't install anything system-wide.

• An NVIDIA GPU. 6 GB VRAM is the realistic floor for SDXL, 8–12 GB is comfortable, 16 GB lets you run Flux. AMD and Intel work through other forks, but this guide is CUDA.
• Disk space. The WebUI is ~8 GB extracted. Every SDXL checkpoint you download is another 6.5 GB, so plan for 40 GB if you like collecting.
• 7-Zip or WinRAR to unpack the .7z archive.
• A matching CUDA build. The common cu124_torch24 package targets RTX 20–40 series. If you're on a 50-series (Blackwell) card, that build will fail with "no kernel image is available" — you need a cu128 or newer package. Run update.bat first; if it still errors, grab the newest release from the repo instead of the older pinned archive.

Install Forge

1. Download the portable build

   Grab webui_forge_cu124_torch24.7z from the Forge WebUI releases page. It's a one-click package — Python, PyTorch and CUDA are all inside.

2. Extract it

   Right-click the archive → Extract All, and pick a folder you'll remember. Keep it on a fast drive with room to grow, and avoid paths with spaces or non-English characters — some extensions still choke on those.

3. Run update.bat

   Do this before the first launch. It pulls the current code over the packaged snapshot, which is usually months old. Let it finish and close on its own.

4. Find a model

   Head to Civitai or Hugging Face. Filter by base model so you know what you're getting — SD 1.5, SDXL, Pony, Illustrious and Flux are not interchangeable, and a LoRA only works with the base it was trained on. Check the license on the model page too: plenty are non-commercial, and a few forbid redistribution of merges.

5. Drop the files in place

   Checkpoints and LoRA live in different folders:

   Checkpoints webui\models\Stable-diffusion

   LoRA webui\models\Lora

   VAE webui\models\VAE

   Embeddings webui\embeddings

   Upscalers webui\models\ESRGAN

   All of these sit inside your extracted folder, e.g. webui_forge_cu124_torch24\webui\models\Stable-diffusion.

6. Launch with run.bat

   A console window opens, then your browser lands on http://127.0.0.1:7860. The first start is slow — it's compiling and caching. Leave the console open; closing it kills the server.

Models, LoRA & VAE

Four different things get called "models" online, and mixing them up is the single most common reason a fresh install produces garbage.

Base families, quickly

Writing prompts

The positive box describes what you want; the negative box describes what to steer away from. Words near the front carry more weight, and the model reads in chunks of 75 tokens.

Different checkpoint families want different phrasing. SDXL photo models like natural sentences; anime models trained on booru data want comma-separated tags. Read the model's own description page — the author almost always posts their recommended settings, and those beat anything generic.

Every setting explained

This is the part the original PDF skipped. Here's what each control in the txt2img panel is doing under the hood.

Sampling methodsampler

The algorithm that walks the image from pure noise to a finished picture. Different samplers reach different images from the same seed. Two broad types: ancestral ones (names ending in a, or containing SDE) inject fresh noise every step, so they keep changing and never fully settle; deterministic ones converge and stop moving after enough steps. Full table below.

Schedule typenoise schedule

How much noise gets removed at each step. Karras takes big early steps and tiny late ones, which usually looks best and is the safe default. Exponential softens output. SGM Uniform is what SDXL-turbo and lightning models expect. Simple / Normal are the plain linear versions.

Sampling steps20–30 typical

How many denoising passes to run. More steps means more time, not automatically more quality — most samplers stop improving around 25–30. Under ~12 you get mush, unless you're on a Turbo/Lightning/LCM model, which are distilled to work in 4–8. Doubling steps roughly doubles generation time.

CFG scaleclassifier-free guidance

How hard the model is pushed to obey your prompt. Low (1–4) is loose and creative but ignores you; high (12+) burns contrast, blows out colour and produces stiff, over-saturated images. 5–7 is the working range for SD 1.5 and SDXL, 2–3 for Turbo/Lightning. Flux uses a separate Distilled CFG slider — 3.5 is its normal value.

Width & Heightresolution

Stay near the model's native pixel count. SD 1.5 wants ~512×512, SDXL and Flux want ~1024×1024. Generating far above native gives you duplicated heads and stretched bodies — the fix is to generate small and use Hires fix, not to type a bigger number. Useful SDXL ratios: 1024×1024 (1:1), 832×1216 (2:3 portrait), 1216×832 (3:2 landscape), 1344×768 (16:9). Keep dimensions multiples of 64.

Seed-1 = random

The number that generates the starting noise. Same seed + same prompt + same settings + same model = the exact same image, every time. That's what makes it useful: leave it at -1 while you're exploring, then click the recycle icon to lock the seed of an image you liked and tweak one thing at a time. A seed is meaningless across different models or resolutions — it won't transfer.

Variation seed & strengthunder "Extra"

Blends a second seed into the first. At 0.1–0.3 strength you get siblings of an image you already like — same composition, different details. Perfect when a generation is 90% right.

Batch count vs Batch sizethroughput

Count runs N generations one after another — costs time, not VRAM. Size generates N at once in parallel — much faster per image but multiplies VRAM use. On 8 GB keep size at 1 and raise count instead.

Hires. fixthe quality button

Generates at native size, upscales the result, then re-runs a partial diffusion pass over it so real detail is invented rather than interpolated. Three controls matter:

• Upscaler — 4x-UltraSharp or 4x_NMKD-Siax for photos, R-ESRGAN 4x+ Anime6B for illustration, Latent for a softer painted look.
• Upscale by — 1.5–2.0. Higher costs VRAM fast.
• Denoising strength — how much freedom the second pass gets. 0.3–0.45 adds detail while keeping the picture; above 0.6 it starts redrawing and re-introducing duplicate limbs.

CLIP skip1 or 2

Stops reading the text encoder one layer early. SD 1.5 anime models (anything descended from NovelAI) expect 2. Photoreal SD 1.5, SDXL and Flux expect 1. Wrong value = flat, slightly-off results.

Refineroptional

Hands the last portion of the steps to a second checkpoint. The switch-at value (0.8 = final 20%) decides when. Mostly unnecessary with modern merged SDXL models — try it, but don't assume it helps.

Sampler reference

Twenty-odd options in the dropdown, four or five worth using. Filter the list if you're looking for something specific.

No sampler matches that.

Speed & VRAM

Forge's big advantage is its memory manager. Before you touch launch flags, use the GPU Weights slider at the top of the UI — it sets how much VRAM is reserved for the model, and leaving 1–2 GB free for the generation itself usually beats any command-line tweak.

For the rest, edit webui-user.bat (or run.bat) and add flags to the COMMANDLINE_ARGS line:

In Windows, also turn off the NVIDIA driver's "CUDA — Sysmem Fallback Policy" for Python if you'd rather see a clean out-of-memory error than have a generation silently crawl at 10× slower while spilling into system RAM.

img2img & inpainting

Same engine, different starting point: instead of pure noise, it starts from your image. One slider dominates everything here.

For inpainting, mask the region you want redrawn and set Only masked so the model works at full resolution on that patch instead of a few dozen pixels. Mask blur around 4–8 hides the seam. This is how you fix hands without rerolling a whole image.

Troubleshooting