This model has previously been named AnimagineXL Realistic or Stylistic and Make Art Not Goon.What makes this model special? I prioritized prompt following and composition, making it excellent for art. This makes it excellent at NSFW as a consequence, not by over training on pr0n at the cost of general knowledge, like many other models.How did I do this? I took the under-appreciated Animagine XL <a target="_blank" rel="ugc" href="https://civarchive.com/models/1188071?modelVersionId=1337429">v4.0</a> because it was well trained from scratch on SDXL 1.0 Base for character knowledge up to Jan 2025. The prompt knowledge of the clip + model is something really special. With the right tags, it is rather artistic and magical with lighting, shadows and composition. <a target="_blank" rel="ugc" href="https://cagliostrolab.net">CagliostroLab</a> have a <a target="_blank" rel="ugc" href="https://discord.gg/cqh9tZgbGc">discord</a> too. I didn't use "Opt" or "Zero" in the mix, as I think they broke the magic with those.I think the style and prompt following peaked in v5, and I pushed too hard to improve realism in v6, v7, v8. I also focused too much on integrating DMD weights to get faster generations with all samplers - both normal samplers and LCM samplers.<h3 id="v9_wip-update">v9_wip update</h3>It's a work in progress, using the animagine clip with illustrious weights, but maintaining prompt following. For general photorealistic generation, end the prompt with explicit or safe followed by:sharp focus, depth of field, fine hair strands texture, photo \(medium\), photorealistic, realistic, masterpiece, high score, great score, absurdresAnd negative:ugly, blurry, lowres, worst quality, low score, bad score, text, sketch, drawing,<h3 id="previous-versions">Previous versions</h3>Check out the workflow in the "Training Data" download! The v8 model is basically tuned for ddpm at cfg 1, 20 steps. However, Euler a works almost as well, and LCM is particularly good as a first pass for architecture, strong shapes, followed by a ddpm pass to add more detail.Many of the early versions are designed for low CFG. If you want to use cfg &gt; 1 then add the cfg norm node in comfyui to stop burn!For characters, the prompt format is:1girl, name, series, rating, stuff, ... , stuff, ..., stuff, ... , masterpiece, high score, great score, absurdres1girl, remilia scarlet, touhou, safe, riding a horse, sunset, masterpiece, high score, great score, absurdresPrompting guide for Animagine: <a target="_blank" rel="ugc" href="https://cagliostrolab.net/posts/optimizing-animagine-xl-40-in-depth-guideline-and-update">https://cagliostrolab.net/posts/optimizing-animagine-xl-40-in-depth-guideline-and-update</a>You don't need to add masterpiece, high score, great score, absurdres to the end of the prompt - it is tuned to work without it. Only add it if you want more anime style.<h2 id="version-history">Version History</h2>v9 - ⭐⭐⭐ - It's almost an entirely new model. The clip encoder is the same, and the clip attention layers come from v5, but many of the weights are 50% or more from realistic illustrious models. Photorealism is better, prompt following is ok, hands and eyes are a little better. v8 - ⭐⭐ - Heavily reworked the original Animagine 4.0 to maintain dynamic range in lighting and shadow, knowledge and prompt following, and vastly improved photo-realism.v7 - ⭐⭐ - Squeezed a little more out of Animagine by optimizing specific booru tags and darkness.v6 - ⭐⭐ - Quite a leap in prompt following, realism, and surrealist art.v5 - ⭐⭐⭐ - Improved prompt following, shadow and light dynamic range, better artist style flexibility, and good anatomy and acceptable hands. It's still a bit more horny than I'd like but it's difficult when all the realism models excel at XXX.v4 - ⭐ - An improvement in anatomy, realism, highlights and shadows. It also follows closely to Animagine XL 4.0. Unfortunately it became more horny, use "safe" in the positive and (nsfw, explicit) in the negative.v3 - ⭐ - I focused on improving the prompt following from the base model, face expressions and only "good enough" photo-reality. Works with wider range of samplers, steps and cfg.v2 - more photorealism but less authentic to Animagine and broke lighting and expressions. v1 - quite authentic to Animagine, but noisy except with specific samplers.<h2 id="license">License</h2>This model adopts the original <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md">CreativeML Open RAIL++-M License</a> from Stability AI without any modifications or additional restrictions. The license terms remain exactly as specified in the original SDXL license, which includes:<ul><li>✅ Permitted: Commercial use, modifications, distributions, private use</li><li>❌ Prohibited: Illegal activities, harmful content generation, discrimination, exploitation</li><li>⚠️ Requirements: Include license copy, state changes, preserve notices</li><li>📝 Warranty: Provided "AS IS" without warranties</li></ul>Please refer to the <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md">original SDXL license</a> for the complete and authoritative terms and conditions.

AnimagineXL - Art not Goon!

The clip l + clip g are identical to Animagine XL 4.0. The VAE is sd_xl_base VAE.The first few layers are 100% Animagine XL 4.0, but it tapers down to ~50-70% in later layers.Trigger words are required for photorealism.It took &gt;100 merges to get to this point, so I will have to track back through my versions to figure out the exact recipe.

Files with SHA256: 0089b8902276decf83d7fb3a2f97644a646cf892dc39d7f981029288fe5ce38a