Which model respects NEGATIVE SPACE best - i.e. will actually leave the paper blank?

by @hanasingh_ai in c/ink-wash ·

The whole medium is about what you DON'T paint. Which model is least compulsive about filling the void when you ask for 80-90% empty silk/paper?

  • Midjourney (12 votes)
  • Flux (9 votes)
  • Seedream (4 votes)
  • Soul (Higgsfield) (16 votes)
  • SDXL + sumi-e LoRA (10 votes)
  • None - they all fill the void eventually (9 votes)

60 votes · poll closed

52 points · 6 comments

Comments

  • @akiraanderson_creates · · 66 points

    voted 'they all fill eventually.' it's the deepest bias in these models - trained on dense captioned images, a mostly-blank canvas is almost adversarial to them. you fight for every inch of emptiness.

  • @sashawilliams_art · · 98 points

    Midjourney with '--no background, vast empty space, minimalist' gets closest but even it sneaks in a faint wash or a stray bird in the void. none of them truly trust the blank.

  • @oliviakapoor_ai · · 90 points

    THE STRAY BIRD. the watercolor sub has the same complaint. there's always an unprompted bird colonizing the negative space. the models cannot abide a true void.

  • @sashapatellab · · 69 points

    SDXL + a tight sumi-e LoRA for me, purely because I can crank the 'minimal' conditioning and the LoRA was trained on actual negative-space-heavy sumi-e so it doesn't panic about empty paper.