<h2 id="description-vdsci7wso">Description</h2>This model is a custom fine-tuned variant based on the Kohaku-XL-Zeta pretrained foundation <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/Kohaku-XL-Zeta">Kohaku-XL-Zeta</a>. Kohaku-XL-Zeta itself is a "raw" base model trained for 1 epoch on 8+ million Danbooru(mostly) images , using 4x NVIDIA 3090 GPUs! Kohaku serves as a flexible starting point for creative adaptations.To enhance encoder stability and inject cross-domain knowledge beyond Danbooru-specific features, the model was merged with ColorfulXL using cosine dissimilarity weighting (0.25 blend ratio). This integration aims to broaden the model’s understanding of natural language and artistic concepts beyond typical Danbooru tagging conventions.Post-merge stabilization involved 6 epochs at 2e-6 learning rate, followed by ongoing fine-tuning at 9e-7 learning rate to refine details. The closest publicly available fine-tune of this lineage is Illustrous, though it uses an earlier Kohaku version with weaker text comprehension. This variant leverages the improved Kohaku-Colorful hybrid (KC), prioritizing non-realistic art generation and creative flexibility over photorealism.Key Notes :<ul><li>Not optimized for realism; best suited for anime/artistic styles.</li><li>Ideal for users seeking a customizable foundation for niche art generation or further fine-tuning experiments.</li></ul><h2 id="donations-ogdo78pp8">Donations</h2>Please contact with us if you may provide some GPU's or money on trainingDOGE: DEw2DR8C7BnF8GgcrfTzUjSnGkuMeJhg83BTC: 3JHv9Hb8kEW8zMAccdgCdZGfrHeMhH1rpN<h2 id="contacts-050x8h05b">Contacts</h2><a target="_blank" rel="ugc" href="https://t.me/recoilme">recoilme</a>

next train epoch<a target="_blank" rel="ugc" href="https://wandb.ai/recoilme/finetuning/runs/5zdqm2f3/workspace?nw=nwuserrecoilme">https://wandb.ai/recoilme/finetuning/runs/5zdqm2f3</a>

kc_v15.safetensors

v1.5

Files with SHA256: 431c69347b9630f866284ea6e42091639bf70da49be2c88a66e5b523fbe51402