Currently, I’m continuing my experiments with high-resolution training using Illustrious. Why is this important? Despite Rouwei (based on Noobai) or even standard Noobai having far more images "under the hood," I often train test LoRAs with complex anatomy. Through this, I’ve noticed that the older Illustrious 0.1 outperforms Noobai in intricate anatomical scenes. There could be many reasons for this, but I believe the key factor is the training approach.

The creator of Illustrious—once sensational, now widely criticized (due to its fundraising policies)—released versions 1.0 and 1.1, which are plagued by a huge number of issues. However, before v2.0 drops, I aim to solve these problems, just as I did with Noobai (which was also problematic initially, but I managed to overcome its training challenges).

Where am I going with this? Right now, I’m testing a ton of approaches and studying the Prodigy optimizer, as it starts failing on Illustrious 1.1. The issue lies with the model, not the optimizer—but both need addressing to stay current. Ultimately, I’ll need to decide whether to stick with Rouwei or fully transition to the new Illustrious and its 1536×1536 resolution, which is crucial for both the community and growth as a creator.

On a side note: Pony v7.0 seems to have released (though not on Civitai, and it’s paywalled)—I’d love to explore that too. Despite some stagnation, image models keep evolving, but training adaptation grows increasingly tricky. I hope that, through collective effort, we’ll keep pushing forward.

P.S. That was a lot of rambling—hope at least someone reads it. :3