Hosioka
Models by this creator
ðŸ§
Baka-Diffusion
93
Baka-Diffusion is a latent diffusion model that has been fine-tuned and modified to push the limits of Stable Diffusion 1.x models. It uses the Danbooru tagging system and is designed to be compatible with various LoRA and LyCORIS models. The model is available in two variants - Baka-Diffusion[General] and Baka-Diffusion[S3D]. The Baka-Diffusion[General] variant was created as a "blank canvas" model, aiming to be compatible with most LoRA/LyCORIS models while maintaining coherency and outperforming the [S3D] variant. It uses various inference tricks to improve issues like color burn and stability at higher CFG scales. The Baka-Diffusion[S3D] variant is designed to bring a subtle 3D textured look and mimic natural lighting, diverging from the typical anime-style lighting. It works well with low rank networks like LoRA and LyCORIS, and is optimized for higher resolutions like 600x896. Model inputs and outputs Inputs Textual prompts**: The model accepts text prompts that describe the desired image, using the Danbooru tagging system. Negative prompts**: The model also accepts negative prompts to exclude certain undesirable elements from the generated image. Outputs Images**: The model generates high-quality anime-style images based on the provided textual prompts. Capabilities The Baka-Diffusion model excels at generating detailed, coherent anime-style images. It is particularly well-suited for creating characters and scenes with a natural, 3D-like appearance. The model's compatibility with LoRA and LyCORIS models allows for further customization and style mixing. What can I use it for? Baka-Diffusion can be used as a powerful tool for creating anime-inspired artwork and illustrations. Its versatility makes it suitable for a wide range of projects, from character design to background creation. The model's ability to generate images with a subtle 3D effect can be particularly useful for creating immersive and visually engaging scenes. Things to try One interesting aspect of Baka-Diffusion is the use of inference tricks, such as leveraging textual inversion, to improve the model's performance and coherency. Experimenting with different textual inversion models or creating your own can be a great way to explore the capabilities of this AI system. Additionally, combining Baka-Diffusion with other LoRA or LyCORIS models can lead to unique and unexpected results, allowing you to blend styles and create truly distinctive artwork.
Updated 5/28/2024
🌀
AniReal
75
AniReal is an experimental text-to-image diffusion model created by maintainer Hosioka. It is a predecessor to the newer Baka-Diffusion model, which aims to push the limits of Stable Diffusion 1.x-based models through techniques like U-Net block merging. The goal with AniReal was to create a "blank canvas" model that could serve as a base for compatibility with various LoRA and LyCORIS models, while providing improved coherency and stability compared to previous iterations. This model diverges from the typical anime-style lighting seen in many Baka-Diffusion variants, instead focusing on a more natural appearance. Model inputs and outputs Inputs Text Prompts**: AniReal accepts natural language text prompts to guide the image generation process. Outputs Images**: The model generates high-quality, detail-oriented images based on the provided text prompts. The output images exhibit a refined, naturalistic aesthetic rather than a purely anime style. Capabilities AniReal demonstrates improved capabilities over earlier Baka-Diffusion models, including more coherent and stable generations, particularly at higher CFG scale values. The model is also designed to work well with low-rank networks like LoRA and LyCORIS, ensuring compatibility and versatility. What can I use it for? The AniReal model can be a useful starting point for a variety of text-to-image projects, from generating detailed anime-inspired artwork to creating naturalistic illustrations. Its flexibility and compatibility with additional models make it a valuable tool for creative work. Things to try One key recommendation from Hosioka is to use a lightweight negative textual inversion, such as SimpleNegativeV1 by Aikimi, to further improve the model's coherency without sacrificing style. The maintainer also suggests experimenting with the FreeU preset, which can help steer the model towards aesthetically pleasing generations.
Updated 5/28/2024