L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

demofusion

demofusion

Configure and troubleshoot the extension's DemoFusion panel for staged multi-scale upscaling, local/global windows, random jitter, mixture mode, noise inversion, and ControlNet/StableSR interop.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: demofusion description: "Configure and troubleshoot the extension's DemoFusion panel for staged multi-scale upscaling, local/global windows, random jitter, mixture mode, noise inversion, and ControlNet/StableSR interop." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

DemoFusion

Use this sub-skill for the extension's separate DemoFusion panel. DemoFusion performs staged multi-scale denoising/upscaling with local tiled windows and global views. It is not the same route as the Tiled Diffusion panel.

When to read

Read this when the user asks about:

  • enabling DemoFusion in AUTOMATIC1111;
  • staged x2/x3/x4 upscale behavior;
  • latent window size, overlap, local/global batch sizes;
  • random jitter, mixture mode, Gaussian filter, cosine scales, or sigma;
  • DemoFusion with img2img, keep-input-size, or noise inversion;
  • why the panel says not to open DemoFusion with Tiled Diffusion;
  • DemoFusion UniPC incompatibility, speed, OOM, or image-size infotext behavior.

For normal Tiled Diffusion/MultiDiffusion/Mixture of Diffusers, route to tiled-diffusion. For VAE encode/decode OOM, route to tiled-vae.

Core operating facts

  • The panel title is DemoFusion.
  • The UI text warns: do not open DemoFusion with Tiled Diffusion.
  • The only method choice is DemoFusion.
  • DemoFusion asserts that it is not compatible with UniPC.
  • Scale factor is used as an integer staged upscale count.
  • Window size defaults to 128 and is bounded by the latent canvas size.
  • Window overlap defaults to 64.
  • Local tile batch size and global window batch size both default to 4.
  • Random jitter defaults on.
  • Cosine scale defaults are 3, 1, and 1; sigma defaults to 0.6.
  • Txt2img exposes denoising strength for the substage with default 0.85.

References

  • Workflows: basic txt2img/img2img DemoFusion setup, staged upscale controls, and parameter meanings.
  • Troubleshooting: conflict with Tiled Diffusion, sampler incompatibility, stochastic jitter, OOM, and slow staged runs.
Skills enregistrées