StableDiffusionWebUI: High-Resolution Fix (Hires.fix) - Algorithms
Preface
SD generates 512x512 images by default.
Setting the width and height too large on the first generation may cause unexpected issues.
It is recommended to generate at 512x512 first, then upscale later to improve resolution.
There are many methods for high-resolution upscaling. This section introduces the High-Resolution Fix (Hires.fix).
Hires.fix
In Hires.fix, the upscale ratio, width, height, and sampling steps can remain unchanged.
The main variables are the upscaling algorithm and denoising strength.
We keep the checkpoint, prompts, and other settings the same to compare these two factors.
Prompt - Chinese
一个男孩，坐在椅子上，上半身，靠近窗口，蓝色上衣，黑色头发，戴眼镜，面向镜头，左手边是窗子
Prompt - English
A boy sitting on a chair, upper body, close to the window, blue shirt, black hair, wearing glasses, facing the camera, with the window on his left side
Checkpoint
Samaritan 3d Cartoon
Seed
3255625742
Sampler
Euler a
Sampling Steps
25
Denoising Strength
Temporarily fixed at 0.7. The next section will compare denoising strength values.
Algorithm: Latent
Algorithm: Latent (antialiased)
Algorithm: Latent (bicubic)
Algorithm: Latent (bicubic antialiased)
Algorithm: Latent (nearest)
Algorithm: Latent (nearest-exact)
Algorithm: None
Algorithm: Lanczos
Algorithm: Nearest
Algorithm: DAT x2
Algorithm: DAT x3
Algorithm: DAT x4
Algorithm: ESRGAN_4x
Algorithm: LDSR
Algorithm: R-ESRGAN 4x+
Algorithm: R-ESRGAN 4x+ Anime6B
Algorithm: ScuNET GAN
Algorithm: ScuNET PSNR
Algorithm: SwinIR 4x
Summary
In this example, the differences between Hires.fix algorithms are not very significant.
Results may vary in different scenarios, so it is worth experimenting.
Two commonly recommended algorithms online:
R-ESRGAN 4x+: Suitable for realistic scenes
R-ESRGAN 4x+ Anime6B: Suitable for anime scenes
SDWebUI
|—Link: https://sdwebui.ai/?from=vq
|—Authentic Stable Diffusion WebUI V1.10.1~
|—Top-tier 4090 GPU, premium members get exclusive single 4090 card for image generation~
|—Supports Mac client, Windows client, and Web online usage~
|—Innovative instant model transfer feature, instantly use local models~
|—Supports txt2img, img2img, extras, and PNG info
|—Supports CLIP interrogation and DeepBooru interrogation
|—Supports Hires.fix with 21 built-in upscaling models
|—Supports ADetailer with 11 built-in repair models
|—Supports ControlNet with 44 built-in preprocessors and 68 built-in models
|—Supports super-resolution upscaling (Tiled Diffusion, Demo Fusion, Tiled VAE)
|—Supports old photo restoration with 21 built-in upscaling models, plus GFPGAN and CodeFormer restoration models
|—Built-in 2 common Styles, 11 popular Checkpoint models, 7 popular LoRA models, 14 common embeddings
|—Built-in oldsix prompt plugin
|—Built-in 23-lesson beginner tutorial