Anima LoRAs fail to capture artistic style despite accurate character reproduction
Users report Anima-based LoRAs accurately reproduce character anatomy and clothing but consistently fail to replicate artistic style, defaulting to generic aesthetics even after multiple training attempts with varied optimizers.
Stable Diffusion practitioners training custom LoRAs on Anima are hitting a consistent wall: the fine-tunes learn character anatomy and outfit details better than Illustrious, but they can't replicate the source material's artistic style.
One user documented five separate training runs on Anima using the Anima Standalone Trainer, testing both Adafactor and AdamW optimizers across 1,500-step sessions. Each run produced accurate character likenesses and clothing, outperforming Illustrious on those metrics. But the output style remained generic, occasionally adding distorted head proportions or bulkier physiques that weren't present in the reference images. Backgrounds defaulted to flat white or solid colors unless explicitly prompted, while Illustrious LoRAs trained on identical datasets preserved the visual tone and environmental context of the training set.
The pattern held across multiple characters and persisted after Anima's full release replaced the preview checkpoint. Illustrious LoRAs trained with the same hyperparameters and image sets consistently matched the reference aesthetic, avoided anatomical drift, and carried forward background atmosphere without extra prompting. The divergence suggests Anima's base model may encode character features more strongly than stylistic attributes during fine-tuning, leaving LoRA adapters with less room to steer the visual treatment.