Janus-Pro — decoupled input encoders, shared transformer body
input: image
source of visual signal
understanding: describe
generation: condition on
SigLIP encoder
understanding path
semantic features
VQ encoder
generation path
reconstruction codes
shared transformer
one body, one weight set
init from DeepSeek-7B
absorbs both encoders
autoregressive decode
text output
NTP, vocab logits
VQA, caption
image VQ -> pixels
emit VQ tokens
decoder -> pixels
routing tag picks encoder and output head
<understand> image -> SigLIP -> body -> text
<generate> text -> body -> VQ tokens -> pixels
Janus-Pro data + scale scoreboard
axis
Janus (Oct 2024)
Janus-Pro (Jan 2025)
delta
model params
1.3B
7B
5.4x
stage-2 data
26M pairs
72M pairs
+176%
MMMU
30.5
60.3
+29.8
GenEval
0.61
0.80 (beats DALL-E 3)
+0.19