Auto: the model picks its own detail level per image — flat areas
merge to few tokens, detailed areas keep theirs, at a steady quality. See τ below the image.
Quality
Higher quality keeps more tokens (lower merge threshold τ).
Solves τ to hit this budget. Slower (re-encodes a few times).
🧪 Experiment Lab
Upload an image, then run the 1m model at every preset head-to-head.
Preset
Output
Tokens used
τ
PSNR
Match
bpp
Efficiency
Grade
Time*
Efficiency = PSNR ÷ bits-per-pixel — quality kept per token spent
(higher is better), graded on this image. Match is structural similarity (SSIM) to the original.
Every row is the same 1m model at a different token budget. Time* depends on your hardware
and is shown for reference only — it is not part of the score.