Emu3 — one model, one loss, three roles
single decoder-only transformer with shared vocabulary
Emu3-Gen
text -> image tokens
512x512 = 4096 tokens
CFG gamma 3-7
temperature 0.8
matches SDXL on FID
no diffusion schedule
no CLIP loss
Emu3-Chat
image -> text
VQA + captioning
matches LLaVA-1.6
same backbone
VQAv2 75.1
unified loss unlocks
perception + gen
Emu3-Stage2
text -> video tokens
4s @ 8fps, 3D VQ
4x4x4 spatiotemporal
patch quantization
competitive FVD
same shared vocab
extends to 10s at scale
Emu3 vs diffusion: the 2026 trade-off
Emu3 (discrete tokens, NTP)
+ one model for gen + perception
+ one training loss
+ tokens extend to any modality
- slow inference (2 min / 512x512)
- tokenizer caps quality
best for research, unified models
Diffusion (SDXL, SD3, Flux)
+ fast inference (2-5s / 512x512)
+ continuous latent = higher fidelity
+ mature tooling + LoRAs
- no perception in same model
- separate text encoder needed
best for image-only production