Illustrious XL 2.0: Increased robustness for better fine-tuning capabilities

For the v2.0 base + aesthetic model’s technical details, check out: Illustrious XL Tech Blog

Stay in the loop with our latest releases and updates at: illustrious-xl.ai

In this version, the dataset has increased drastically, including major focus on animations & natural language focus – with updated dataset, until 2024-08. The model has performed best finetuning stability in our internal tests – with highest TrueSkill rating(among v1-v2) in external demo evaluation.

We want to note that we consider the compatibility & finetuning capability the most important feature of our baseline model. The model itself is not trained with aesthetic set / and we tried our best to debias the potential biases in model, for better overall performance.

Thus, we release ‘untuned’ base version, which should work as better merging / training bases. With release of the models, we also showcase example LoRA adapters, which is widely compatible across versions with recent character knowledges.

The model itself is highly compatible with natural language sentences – it is far more robust, it is less likely to generate multiple views or nonsense outputs.

V2.0 supports 512~1536 resolutions. It is trained max 1:10 (or, 10:1) aspect ratio, with width / height multiples of 32. The models are capable for extreme high resolution generation more stably, with higher denoising strength, which is important for stable detail generations. We are also analyzing how to tweak the sampler schedule for better detail generations – we will try our best to showcase the results.

Finally, we note the definite future plan for the model releases.

We have decided to make model release sequentially, in three stage manner:

Exclusive
Early access
Open source

We will also continue to train & release the future model series, including DiT based models.

Illustrious XL - v2.0 - BASE