2024-12-01 Update: Added inpainting version.Somewhere around 2.71̶̦̗̋́̓̍̓8̵̰͚̪͂̽͋́̂̃͛̕ D. <a target="_blank" rel="ugc" href="https://civarchive.com/models/549840">Recursed Canvas</a> is a less realistic, fantasy-oriented Stable Diffusion 1.5 model. This model is tailored for creating imaginative and otherworldly art, embracing the surreal and fantastical elements that escape the constraints of reality. Its tuning emphasizes vibrant, dreamlike visuals and creative abstraction, with a bias towards Western art styles. This model is meant to be much more creative and artistic than the other model it is meant to play off of, <a target="_blank" rel="ugc" href="https://civarchive.com/models/549644">Aberrated Perceptron</a>. I blend the two using <a target="_blank" rel="ugc" href="https://civarchive.com/models/549840">Recursed Canvas</a> as the base using the various methods provided by A1111, such as the refiner, <a target="_blank" rel="ugc" href="https://github.com/w-e-w/sd-webui-hires-fix-tweaks">hires</a>, and in-painting with tools like <a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">Adetailer</a>. Using <a target="_blank" rel="ugc" href="https://civarchive.com/models/549840">Recursed Canvas</a> for the first 20-95% has had interesting results. Example <a target="_blank" rel="ugc" href="https://civarchive.com/posts/4738832">here</a>. This was mostly an experiment focused on understanding the strengths, weaknesses, and limitations of realistic image generation. These models serves as my own lazy-tuned benchmark for pushing the boundaries of realism in AI-generated art within the limitations of SD 1.5.I'm not an artist, I'm not creative. I can make an LLM be creative for me. When testing this model, I used ChatGPT to create my own tool of uninspired laziness: <a target="_blank" rel="ugc" href="https://chatgpt.com/g/g-XasBw0Gci-fantasy-art-prompt-generator">Fantasty Art Prompt Generator</a>Suggested Resources below has the list of sources for the previous versions of this model as far back as I can see. Models after 1.7 have <a target="_blank" rel="ugc" href="https://civarchive.com/models/23906/kl-f8-anime2-vae">kl-f8-anime2 VAE</a> embedded.

Recursed Canvas

Identified an issue with 1.8 where artifacts were being introduced at the lower edge of generated images. Also identified an issue with CFG &gt;7. I merged with <a target="_blank" rel="ugc" href="https://civitai.com/models/193551?modelVersionId=293271">RaekanaMix 2.5D</a> with different MBW weights in an attempt to address these issues. I'll be using v2.0 for the version since this version includes a fairly major change instead of a bugfix version (e.g. v1.8.1). Merge metadata is included in the checkpoint info.

Files with SHA256: 1647882a4ea3ab5a010381c12e557dc0957608b19d535be59ae072a4bf58cc39