Try size: 512 x 768 (SD 1.5) / 768 x 1024 (Pony and ILXL and)

Weights: 0.8~1

Clip skip: 2

Training material from bing

Pictures from Calicomix 7.5 and CalicomixPonyXL_v10 and CalicomixEroill_v30 and

Note: Fingers and dolphin are not very stable. (SD 1.5)

Note: The shark is more interesting as example images, so provide trigger words. The other prompts for dolphins, please refer to pictures. (Pony)

Note: The orca is more interesting as example images, so provide trigger words. The other prompts for dolphins, please refer to pictures. (ILXL)

Note: The orca is more interesting as example images, so provide trigger words. I believe the new parameters can make it more stable. For OOO, the "Debiased Estimation loss and AdamW and constant_with_warmup" are better. (OOO)

Note: I used Animagine XL 4.0 Zero training Lora, but the training results still disappointed me. It is undeniable that it is an excellent model, maybe I will have the opportunity to try its next version in the future. (ANI4)

(Actually, the name aquarium could cover a wider range. I hope you don’t mind whether it is a dolphin. Dolphins account for about 70% to 80% of the training material).

(For SD 1.5 vesion file size 37.27 MB are all trained on pixai in the past, the parameters are a mystery, so the results are either very good or very bad).