Use <a target="_blank" rel="ugc" href="https://civarchive.com/models/2560840/anima-turbo-lora">https://civarchive.com/models/2560840/anima-turbo-lora</a> with the model if you want similar output. Use 8-12 steps, more steps usually giving more crips images. You can probably get the same without but I have no idea about the negatives you would need. Heun + beta is so far the best combination of sampler and scheduler I have found, creating the most lifelike images. This model is a modest realistic finetune with Anima, which seems to be surprisingly good at learning the photographic rendition. It has been trained for 50 epochs on same dataset as LoZ, balanced slightly. At the moment that means roughly 40k steps at BS 16 and will continue for another 50, culling part of the dataset for the next 50 epochs and see what happens.For prompts, do not use the quality tags of Anima, but you can use tag prompting or natural language.After full training was run, the model seems more competent on variety of sampler + scheduler combinations, but still falls apart in more complex scenes. I would say it is a fairly competent 1girl generator for the amount of effort put in though. Whatever was left of text producing capacity is mostly nuked out and I think the model defaults to photographic generation now.

Panima

Half of the expected 100 epochs trained

Files with SHA256: b4d84aabcc5641b1b9df0e12fcc664ad71a52cbad58cc98ccb75f2f4799874f3