Boosts sharpness and clears up details at higher resolutions. Also enhances direct generation at higher resolutions. The sample images were all directly generated at that resolution.
Can also be used to clear up the image in the 'Hirez-fix' step.
Limitations:
The sharpness bias might not be suitable for all styles, try reducing the step count.
There is a bit of saturation and colour bias, so you might have to post-process.
Description
Initial version. Trained on 1024, 1536, and 1792 multires dataset.
FAQ
Comments (3)
Could you share a bit of info on how this was created?
Process is pretty straightforward. Collected good quality very high resolution illustrations, tagged them with anime caption model, and trained a DoRA with OneTrainer.
penguin235Oct 1, 2026
@latent_space_dreams thanks!