Epoch 1 warmup ckpt for training LoRA. Not intended to gen images.v1.5 is based on Anima base v1<hr />What's the model for?Removes high 2d art bias in base anima. Prevents LoRAs learning same high counter bias. So you can stack more LoRAs without accumulated bias breaking the model.A simplified example:The base anima has high bias towards flat 2d art without textures and details. And you have 3 digital art datasets, which have much more details than base anima. If you training 3 LoRAs from those datasets on base anima. Aside of learning concepts, every LoRA will also learn how to counter the bias to add more details. If you stack those 3 LoRAs, all of them want to add more details, and there will be artifacts etc.By training the base model with all those datasets first, you get an intermediate ckpt with more details. If you training LoRA on this ckpt, the LoRA will not learn the counter bias motioned above.However, in practice it's more complicated.Do I need this model?It heavily depends on the your training data.

Files with SHA256: 72ee982a56a60541a29fbb0d567e556f875d0f79b79febf113699faf2eec08f9