Cartethyia/Fleurdelys | Wuthering Waves (NAI/IL/Pony)

<ul><li>Training time: 10h40m</li><li>11369 images: 10208 "3d", 810 fan art, 8370 "cartethyia", 2258 "fleurdelys", 275 "leviathan", 817 "male rover", 487 "female rover", 129 "iuno", 104 "phoebe", 1843 "video artifacts"</li><li>OneTrainer</li><li>32/32 LoRA, dropout=0.2, fp32 training precision, tf32 mixed precision, fp16 output</li><li>Optimizer: ProdigyPlusScheduleFree(betas=(0.9,0.99), weight_decay=0, weight_decay_by_lr=False, bias_correction=False, d0=1e-4, prodigy_steps=0, eps=1e-8, factored_fp32=True, use_speed=True, split_groups=True, split_groups_mean=False, factored=True, fused_back_pass=True)</li><li>Scheduler: Constant(lr=1)</li><li>batch size: 32, accumulation steps: 1, stopped early at 3000 steps (8.4 epochs), manually evaluated every 100 steps</li><li>no clip grad norm (StableAdamW handles gradient clipping internally)</li><li>Train TE1&amp;TE2, dropout=0.1</li><li>offset noise weight: 0.01, generalized offset noise, uniform timestep distribution</li><li>loss function: MSE, loss weight: debiased estimation loss</li><li>layer filter: attn-mlp</li><li>VRAM: Average was probably around 75 GB with peaks close to 93 GB</li><li>Unet LR quickly settled around 3.26e-5, TE LR remained constant d0=1e-4 (optimizer bug or settings problem?)</li><li>After 3k steps parts of her outfit started appearing on Hakurei Reimu, so I stopped training.</li></ul>

cartethyia-noob-v5.safetensors

Cartethyia/Fleurdelys | Wuthering Waves

v5 [eps]

v4 [Noob eps/v-pred/IL]

v3 [Noob v-pred 1.0]

v3 [Noob eps 1.1]

v3 [IL 0.1]

v3 [Pony]

v2 [Noob v-pred 1.0]

v2 [Noob eps 1.1]

v2 [Illustrious 0.1]

v2 [Pony]

v1-noob-e

v1-pdxl

v1-illust

Files with SHA256: 9b8667a8509d6869b01c8ae8d34abc7792e4779b44096be4e21c16e99c053636