This is an SDXL 1.0-based LoRA and is largely incompatible with ILL. ILL’s architecture has shifted too far from the base 1.0 weights. As an anime-centric model, it struggles with textures derived from real-life photography. Given that remaking it would require excessive resources for data collection and training—and because ILL can already approximate these effects via prompts—I decided against retraining.

Thanks for the feedback; I'm glad to know what kind of issues are popping up.

@Dante3rd I’m glad you like it! However, the biggest issue isn’t the cost, but rather that I’m not sure if the training will be successful. Most of my previous training attempts on ill haven't turned out very well. If it works, I’ll release it for free. Alternatively, have you considered using the Anima model? It’s generally easier to train, and you might even get the results you're looking for without any additional training.

@Dante3rd Is your ComfyUI up to date? Make sure you’ve downloaded the necessary CLIP, VAE, and relevant workflows. I suggest grabbing a workflow from Civitai that suits your taste; get the basic setup working before you start customizing it. The Anima model retains some realistic photographic traits, allowing for more diverse style shifts. I've been using it as the base for my recent training experiments.

@Dante3rd Additionally, LoRAs are only compatible if they were trained on the matching model. When using the Anima model, ensure your LoRA was specifically trained with Anima as its base

@Dante3rd I mostly use the latest ComfyUI build, but be aware that updating can lead to broken nodes or errors in your existing workflows