[HN Gopher] Wurstchen: Fast Diffusion for Image Generation
       ___________________________________________________________________
        
       Wurstchen: Fast Diffusion for Image Generation
        
       Author : cmitsakis
       Score  : 16 points
       Date   : 2023-09-13 19:02 UTC (3 hours ago)
        
 (HTM) web link (huggingface.co)
 (TXT) w3m dump (huggingface.co)
        
       | webmaven wrote:
       | Some cursory (and shallow) testing shows that the model does
       | reasonably well at representing artist styles compared to SDXL,
       | certainly in terms of genre and color palette.
       | 
       | It actually does a bit better in terms of reproducing an artist's
       | typical texture and technique (eg. "Art by Virgil Finlay"
       | reproduces that artists' stippling technique without having to
       | specify anything further like a particular medium).
       | 
       | In terms of artistic composition, results often seem a bit
       | simplified or smoothed compared to SDXL (or SD v2.1 for that
       | matter). Almost cartoonish, and reminiscent of Craiyon or other
       | early models.
       | 
       | Interpolating between artist styles are about on par with SDXL,
       | though the results are somewhat less consistent (that is, a
       | series of "By ARTIST_1 and ARTIST_2" from SDXL will tend to share
       | a common style, and from Wurstchen they will be more diverse).
        
       | cubefox wrote:
       | I wonder how this compares to GAN based text-to-image models like
       | StyleGAN-T. If I remember correctly, GAN models mainly shine at
       | very fast inference, but the same may not be true for training.
       | Also diffusion based models seem to have generally higher
       | quality.
        
       ___________________________________________________________________
       (page generated 2023-09-13 23:02 UTC)