Special thanks to @KH38MT for providing gpt-image-2 access, which made the dataset collection much quicker. Feel free to check out his profile for the loras he published too!<h1 id="quick-version-guide">⚡ Quick Version Guide</h1>v1Default girl = mature type.Exquisite details and beautiful backgrounds.👉 Pick this if you like the mature lady vibe. v2-petiteDefault girl = petite / small body type.Cleaner image with less "noise" (fewer details, simpler backgrounds).👉 Pick this if you want petite girls and prefer a cleaner look. v2.1 ✨Default girl = petite / small body type.Details and background quality on par with v1.👉 Pick this if you want petite girls BUT also want the exquisite details from v1. v3-fusionDefault girl = neutral (not locked to petite or mature).Detail level sits between v1 and v2.👉 Pick this if you want a versatile all-round version and can accept slightly less detail than v1.<h1 id="basic-info">📌 Basic Info</h1>All the training data comes from images I generated using a style I really loved that I rolled from gpt-image-2.Train model: Anima_base_v1.0Trigger word: @gpt2Weight: 0.8 ~ 1.5CFG: 4.5<h3 id="sampler-+-scheduler">Sampler + Scheduler</h3>Best: exp_heun_2_x0_sde + beta (best quality but very slow)Fine: dpmpp_3m_sde + sgm_uniform---<h1 id="version-notes">📝 Version Notes</h1><h2 id="v2.1-(latest)">v2.1 (Latest)</h2>This is released after v3. It's called v2.1 because it's a direct upgrade to v2, not a new direction.I took v1's dataset captions, changed the subject from mature woman to petite girl, and trained a new lora.So basically you get v2's petite body type with v1 level details and backgrounds.If you liked v2 but wished it had more details, this is the one.<h2 id="v3-fusion">v3-fusion</h2>So... v2 did make the characters petite, but it came with a cost.It lost a lot of details. The compositions felt less refined, and the backgrounds became super plain unless you prompt really hard.To fix this, I combined the v1 and v2 datasets together to train v3.Pros 👍:Brings back the "exquisite" details and beautiful effortless backgrounds from v1.Still keeps the default "1girl" output in a smaller, youthful "maiden" style (won't revert back to v1's mature look).Basically v3 is the best of both worlds. You get the details from v1 with the body proportions from v2.<h2 id="v2">v2</h2>v1 leans too much towards mature characters and can't really do "petite" well.To fix this, I generated 16 new petite images using gpt-image-2 to train v2.(Even though I'm not really into petite women myself lol)The v2 dataset is completely different from v1, but the art style stays exactly the same.BTW if you want to use "The Tag That Must Not Be Named" 🥑, v2 works WAY better than v1.<h2 id="v1">v1</h2>Pros 👍: You get beautiful backgrounds without specifying details in your prompt, and no extra detailer needed. <h1 id="v2.1">V2.1</h1><h1><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/491fca73-871d-400e-a89e-f08a5a4b9c97/original=true/491fca73-871d-400e-a89e-f08a5a4b9c97.jpeg" /></h1><h1 id="v3:">v3:</h1><h1><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/17c8f347-615a-47a6-9b3e-1682be4da881/original=true/17c8f347-615a-47a6-9b3e-1682be4da881.jpeg" /></h1><h1 id="v2:">v2:</h1><h1 id="v1:"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d428d8b8-657d-4412-8adf-0fa803dce8d8/original=true/d428d8b8-657d-4412-8adf-0fa803dce8d8.jpeg" />v1: <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7000b5f9-bcd8-4176-820b-4b084d93509a/original=true/7000b5f9-bcd8-4176-820b-4b084d93509a.jpeg" /></h1>

Ps GPT-image-2 style

Settings (Same as v1):<ul><li>Trigger word: <code>@gpt2</code></li><li>Weight: 0.6 ~ 1.5</li><li>Sampler: Dpmpp_3m_sde</li><li>Scheduler: sgm_uniform</li><li>CFG: 4.5</li><li>Base model: Anima_base_v1.0</li></ul>v1 leans too much towards mature characters and can't really do "petite". To fix this, I generated 16 new petite images using gpt-image-2 to train v2. (Even though I'm not really into petite girl myself) The v2 dataset is completely different from v1, but the art style is kept exactly the same. BTW, if you want to use "The Tag That Must Not Be Named" 🥑, v2 works WAY better than v1.

Files with SHA256: 92ce150328e03700f7b825d9170224fd6f7e811bbf672d631297ece5c478ecc7