Show-o — parallel masked-discrete-diffusion image sampling
cosine mask schedule over 8 steps
step 0
step 2
step 4
step 6
step 8
16 masked
12 masked
8 masked
2 masked
all filled
at each step predict all masks in parallel, commit top-K confident
Show-o vs alternatives
Chameleon / Emu3
discrete + NTP
1024 forward passes
~2 min / 512x512
simplest training
tokenizer-capped quality
Show-o / MaskGIT
discrete + masked diff
~16 forward passes
~4-8s / 512x512
single loss, clean
inpainting free
Transfusion / MMDiT
continuous + diffusion
~20 forward passes
~5-10s / 512x512
highest quality
dual-loss to tune
Stable Diffusion
continuous latent
~20 passes
~2-5s / 512x512
specialist model
no VQA/reasoning