What does Clip Skip do and what should I set it to?

Clip Skip controls how many of the final CLIP text-encoder layers are skipped when encoding your prompt. Photoreal SD 1.5 and SDXL models generally want Clip Skip 1. Many anime and NovelAI-derived models are trained expecting Clip Skip 2, which gives cleaner results on those checkpoints. Match it to the model's documentation.

Do I need a separate VAE?

Some checkpoints ship with a baked-in VAE and some do not. If your outputs look washed out, overly gray, or have purple artifacts, loading a dedicated VAE (like the SDXL VAE or vae-ft-mse for SD 1.5) usually fixes it. Set it in Settings under the VAE section so it applies globally.

What is xformers and should I enable it?

xformers is a memory-efficient attention implementation that speeds up generation and lowers VRAM use on NVIDIA cards. Enable it with the --xformers launch flag if your card supports it. On newer PyTorch builds, the built-in SDP attention is an alternative that needs no extra install.

Are these the only settings?

No. Automatic1111 exposes hundreds of options across many tabs, plus extension settings. This guide covers the high-impact settings most people need to get good output, organized by area, rather than documenting every checkbox.

What is the Automatic1111 WebUI Settings Guide?

Guide to Automatic1111 WebUI settings — Stable Diffusion options, VAE, optimization, face restoration, upscaling, and sampling defaults. Filter by category and use case to see what each setting does and a recommended starting value. It runs free in your browser on Gera Tools, with nothing uploaded.

Automatic1111 WebUI Settings Guide

Name: Automatic1111 WebUI Settings Guide
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Automatic1111 WebUI settings guide

The Automatic1111 WebUI is powerful but its Settings tab is a wall of options with little explanation, and a few of them quietly decide whether your images look good. This guide pulls out the high-impact settings — Clip Skip, VAE, optimization flags, face restoration, and upscaling — explains what each does, and gives a recommended starting value tuned to your use case.

How it works

Choose a category to match how the WebUI groups settings, then a use case (photoreal, anime, or speed). The guide shows each relevant setting with a plain explanation and a recommended value for that workflow. Settings that only take effect after a restart are flagged, so you are not left wondering why a change did nothing.

The settings that matter most

Most Automatic1111 users spend hours tweaking samplers and CFG scales before realizing a handful of settings options have a larger impact on output quality than any prompt adjustment.

Clip Skip

Clip Skip controls how many final layers of the CLIP text encoder are skipped when converting your prompt into conditioning. Set it to 1 for SD 1.5 photoreal checkpoints and all SDXL models — these are trained with all layers intact and produce slightly flat or misaligned output at Clip Skip 2. Set it to 2 for most anime-style and NovelAI-derived checkpoints, which were trained expecting a shorter CLIP chain. The wrong Clip Skip does not break images dramatically, but it consistently softens prompt adherence — you will notice keywords not reading as strongly as they should.

VAE (Variational Autoencoder)

The VAE decodes the latent image representation into pixels. If your outputs look washed-out, gray, or have purple/pink artifacts in hair and shadows, the checkpoint shipped without a baked-in VAE and you need to load one. For SD 1.5 models, vae-ft-mse-840000-ema-pruned is the community-standard choice. SDXL has its own dedicated VAE. Set the VAE globally in Settings → Stable Diffusion → SD VAE so every generation uses it without needing to set it per prompt.

Attention optimization

On NVIDIA hardware, enabling xformers with the --xformers launch flag cuts VRAM use substantially and speeds up generation. On newer PyTorch 2.x builds, the built-in SDP (scaled dot-product) attention is an alternative that needs no extra install — enable it via --opt-sdp-attention. Pick one; they do not stack. Both changes require a WebUI restart.

Face restoration

CodeFormer and GFPGAN can sharpen distorted or low-resolution faces, but at full strength they create a distinctive “plastic face” look. Set the CodeFormer visibility to 0.3–0.5 for a subtle enhancement and keep weight at 0.5 for a balance between restoration and fidelity to the original generated face. Higher values override details aggressively.

Upscaling

For hires.fix upscaling, R-ESRGAN 4x+ is the general-purpose choice for photoreal images. R-ESRGAN 4x+ Anime6B is preferable for anime-style output. LDSR is higher quality but much slower. Set the denoising strength in hires.fix to 0.5–0.65: below 0.4 adds almost no detail, above 0.7 substantially changes the composition.

Tips for a clean setup

Set Clip Skip per model family. 1 for photoreal/SDXL, 2 for most anime checkpoints — a wrong value subtly degrades every image.
Load a VAE if colors look off. Washed-out or purple-tinted output almost always means a missing or mismatched VAE.
Enable an attention optimization. xformers or SDP attention is the single biggest free speed and VRAM win on NVIDIA hardware.
Use face restoration sparingly. CodeFormer and GFPGAN fix distorted faces but can plasticize skin; lower the visibility weight (0.3–0.5) rather than running at full.
Remember restart requirements. Launch flags like --xformers and VAE changes take effect on WebUI restart. If a setting change appears to do nothing, restart the process.