NoobAI-XL (NAI-XL) Reviews
The further this gets into training, the more apparent the team working on the model are inexperienced as evidenced by their poor foresight in decision making.
The team started safe and produced stellar results, with each increment of training being better than the last and the experiments they conducted showing considerable improvements with colors (as seen in the V-Pred model). Seeing these results, the team rightfully decided on a V-Pred conversion at the end. This positive trend continued until the release of 0.5 Eps Pred. Afterwards, the team seemingly decides to throw caution to the wind by going against the advice of the papers from the makers of the models of NAIv3 and Illustrious 0.1, as well as repeating the mistakes of flawed models such as Pony v6.
The first of these mistakes was training the text encoder, this decision is highly effective at inducing more visible training results in a shorter amount of time, however the consequence of this decision is that training the text encoder is a destructive process, and is further exacerbated by the fact that the team combined the tagging system of another image board which causes major conflicts in tagging. Apart from the tagging schemes of different image boards conflicting with each other, some artists suffered heavily from the text encoder training process, showing a marked decrease in quality. While further training with a frozen text encoder can help realign the model back to convergence (thus returning the quality), it is a well documented truth that this training will cause loss in knowledge from both the SDXL and Illustrious models that this model is trained on.
Now the team plans to drop artist tags from its dataset for data that is older than 2019. The goal of this is to have artist tags produce higher fidelity results of their more recent, and thus higher quality works. This decision is flawed in many ways, and the fact that this decision made it through the entire team perplexes all my colleagues as well. The first consequence of this decision is that it will make artist tags weaker, to the point that prompting the artist will cause no discernable changes in the output - a result of the artist tag having less data. While artists generally do improve their skills with time, elements of the artist's style is often still visible in their earlier works. Furthermore, with supplementary tagging such as year tags or data range tags (i.e. newest, medial, oldest and such), the artist prompts will be able to reproduce outputs in the likeness of an artist's more recent style. The second issue with Noob's decision to remove artist tags prior to 2019 is that to some users, an artist's older style may be a more desirable result than their newer style. Take for example, a user that wants to prompt a character from a JRPG or visual novel older than 2019 in the likeness their respective games' art style. This would require the artist's style at that moment in time to be promptable for the best effect. Or perhaps a user might find Kantoku's older style to be more to their liking than their newer style, or Ishikei's for that matter. Taking into consideration these consequences, the decision to remove artist tags prior to 2019 is bad at worst, and extremely questionable at best.
I implore the team to take a step back and consider consequences of the choices of other models, and their effects. If the training keeps going in this direction, I fear that this project will be just another lesson in what not do when training a model.
