pkuliyi2015/multidiffusion-upscaler-for-automatic1111
Multidiffusion Upscaler For Automatic1111
Este proyecto es una herramienta de escalado de imágenes (upscaling) y generación de alta resolución mediante IA. Utiliza difusión en mosaico (tiled diffusion) para crear imágenes ultra grandes procesándolas en regiones más pequeñas y superpuestas para evitar fallos de memoria en hardware limitado.
El sistema gestiona la composición espacial mediante prompting regional, que enruta prompts de texto específicos a áreas designadas de una imagen. Mantiene la estabilidad visual y la coherencia global durante el proceso de escalado utilizando inversión de ruido y guía estructural.
Las capacidades adicionales incluyen escalado de detalles en mosaico y optimización de memoria para las etapas de codificación y decodificación del autoencoder variacional. Estas características reducen el pico de memoria requerido para la reconstrucción de imágenes de alta resolución.
Features
- Tiled Processing - Uses tiled processing to split large image generations into overlapping regions, preventing video memory overflow.
- Tiled Decoding - Processes the variational autoencoder stage in small tiles to reduce peak memory during high-resolution reconstruction.
- Tiled Latent Processing - Implements latent space tiling to enable the generation of ultra-large images without memory crashes.
- High-Resolution Synthesis - Generates ultra-large, high-fidelity synthetic images by processing them in smaller tiles to prevent hardware crashes.
- Tiled Synthesis - Creates ultra-large images from scratch using tiled diffusion and memory optimization.
- Noise Inversion Techniques - Provides noise inversion on individual tiles to ensure visual stability and coherence during the upscaling process.
- Guided Tiling - Combines tiled generation with external spatial maps to maintain global composition across multiple local processing tiles.
- Tiled Upscaling - Increases image resolution and detail using a tile-based refinement process to maintain quality at scale.
- VAE Memory Optimizations - Reduces peak memory consumption by processing the VAE encoding and decoding stages in small tiles.
- Composition-Controlled Generators - Employs structural guides and tiled generation to manage the precise layout and composition of high-resolution outputs.
- Regional Prompting - Allows assigning specific prompts to different designated areas of an image to control local composition.
- Detail Refinement - Iteratively processes overlapping image segments to add high-resolution details without memory overflow.
- Consistency Inversion - Performs noise inversion on tiles to maintain visual consistency during high-resolution processing.
