We're a Day-0 launch partner for MiniMax H3, a next-gen open-weights multimodal video model - and it's already in the generator. No waitlist. Open weights land in a few days.

MiniMax H3 is live: open-weights video, generate now

<h1 id="lewdxl-v-pred-finetune">LewdXL - V-Pred Finetune</h1>Anime-focused SDXL checkpoint finetune using V-Prediction, beware this is my first ever finetune.<h2 id="artist-training-dataset-(as-of-v0.2)">🎨 Artist Training Dataset (as of v0.2)</h2>This model is an experimental anime-style finetune trained on curated datasets inspired by the following artists:<ul><li>kuto_(kuroto)</li><li>shiomi_(lowrise)</li><li>deadpurity (AI Artist)</li><li>rolua</li><li>Pluvium_grandis</li><li>murai-renji / range murata</li><li>baifeidaiwang</li><li>chen_bin</li><li>hanozuku</li><li>8rk9</li><li>narue</li><li>akifn</li><li>hatena (nazequestion)</li><li>dor_m</li><li>gagaimo</li><li>sweetonedollar</li><li>popepopo999</li></ul><hr />Model is intended for creative and artistic use. Please refer to the base SDXL license for usage terms.

restarted, different artists and smaller dataset. 28mmalphonse_(white_datura)anversaillesas109asakurafasteroid_illbetabeetdiathorndoribaefajyoborefataaafcjfiorfiz-rotflassofuyumurasakihideousbeinghigashiyama_shouhongchamashuhonnryou_hanaruhwoi_(hw_123)jun_daniellkanzarinkarylnkonya_karasuekuroshiro00luicentmedamayaki27760mocha_(cotton)mokewmonomonomurata_rangerellarurikonshio_(shiofeifei)so_shiosushispinsweetonedollartacshojosoratakagaki_erutokyogensovoid_0wang-xiwashiya0xingzhi_lvyu-y_(824412)zandra

Alpha v0.3

retrained from scratch + 2 extra artists:shiomi_(lowrise)kuto_(kuroto)

Alpha v0.2

Artists:<ul><li>rolua</li><li>deadpurity (AI Artist)</li><li>Pluvium_grandis</li><li>murai-renji / range murata</li><li>baifeidaiwang</li><li>chen_bin</li><li>hanozuku</li><li>8rk9</li><li>narue</li><li>akifn</li><li>hatena (nazequestion)</li><li>dor_m</li><li>gagaimo</li><li>sweetonedollar</li><li>popepopo999</li></ul>⚙️ Training ConfigurationCore Parameters<ul><li>Base Model: Zeronansv9 (Noobai based)/Stable Diffusion XL (SDXL)</li><li>Prediction Mode: V-Prediction (<code>force_v_prediction: true</code>)</li><li>Training Epochs: 50</li><li>Training Steps: 10150</li><li>Batch Size: 8</li><li>Learning Rate: <code>6e-6</code></li><li>Text Encoder LR: <code>2e-6</code></li><li>Resolution: <code>1024x1024</code></li></ul>Optimization Settings<ul><li>Scheduler: 1-cycle cosine</li><li>Warmup Steps: 100</li><li>Min LR Factor: 0.1</li><li>Gradient Accumulation: 1</li><li>Gradient Norm Clipping: 1.0</li><li>Dropout: 0.05</li><li>Precision: BF16</li></ul>Loss Configuration<ul><li>MSE Loss: <code>1.0</code></li><li>MAE Loss: <code>0.1</code></li><li>CosH/VB Loss: <code>0.0</code></li><li>Constant Loss Weight: <code>2.0</code></li></ul>🧠 Infrastructure &amp; Training Setup<ul><li>Trainer Framework: OneTrainer</li><li>Platform: RunPod (Community Cloud)</li><li>Hardware: L40S GPU</li><li>Device: CUDA</li><li>EMA: Enabled (<code>decay: 0.999</code>, <code>update_interval: 5</code>)</li><li>Sync Method: SCP (native file sync)</li><li>Autocast Caching: Enabled</li></ul>🔧 Model ComponentsAll major components were actively trained:<ul><li>UNet: ✅ Trained</li><li>Text Encoders (4x): ✅ Trained</li><li>VAE: ✅ Trained</li><li>Prior: ✅ Trained</li><li>EffNet/Decoder/VQGAN: ✅ Trained</li></ul>Additional Settings<ul><li>Masked Training: Disabled</li><li>Offset Noise Weight: <code>0.05</code></li><li>Perturbation Noise Weight: <code>0.05</code></li><li>Timestep Distribution: Uniform [0.0 → 1.0]</li></ul>📦 Release NotesThis is the initial alpha release (<code>v0.1</code>) of LewdXL, focused on establishing baseline anime style reproduction with V-Prediction architecture. This experimental checkpoint tests core functionality and artistic style transfer capabilities.Expected Improvements in Future Versions:<ul><li>Enhanced prompt adherence and control</li><li>Improved image coherency and consistency</li><li>Greater output diversity and variation</li><li>Refined artist style blending</li><li>More artists, concepts and characters</li></ul>