Emu3-Gen¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
emu3-gen |
BAAI/Emu3-Gen | 40b3bb8 |
Apache-2.0 | Verified |
Catalog: emu3 ·
Model: emu3-gen · Task: t2i
Checkpoint. emu3-gen pins the author repository BAAI/Emu3-Gen to revision
40b3bb89a9cee99ab0b9deb35dead2e50d6a0d72 and the required
BAAI/Emu3-VisionTokenizer to c81f916ad371289e205310a7539255e8a9396488. Both are
Apache-2.0. The architecture reference is the official Emu3 repository at
dbbf9858194d70b8c58293e219ecffe22df0f9c7.
Architecture. Emu3-Gen is an autoregressive multimodal decoder rather than a diffusion model. Its 32 Llama-style blocks use a 4096-wide hidden state, grouped-query attention with 32 query and eight key/value heads, SwiGLU feed-forward layers and rotary positions. Text and image control markers share a 184,622-token vocabulary with 32,768 visual tokens. A separate spatiotemporal VQ model maps the generated grid through a four-channel codebook and conditioned convolutional decoder.
Generation. The backend formats the published BOS, image-start, grid-size and image markers,
then samples visual tokens left to right. Row endings and image/end markers are forced at their
published positions. Classifier-free guidance uses the official positive suffix and negative prompt.
The VQ decoder expands each token by eight pixels and converts its [-1, 1] output to uint8.
Settings. Defaults are 720×720, guidance 3.0, temperature 1.0 and top-k 2048. width and
height must be positive multiples of eight whose product does not exceed 518,400. The backend also
accepts seed, negative_prompt and positive_suffix. Unknown settings are rejected before loading.
Validation and limits. Published layouts cover all 291 generator and 495 vision-tokenizer tensors. Exact random-weight comparisons cover the language model and vision decoder. Real tests compare components and separate seed-0 and seed-42 images through the public pipeline at 24×16; the deliberately small grid retains prompt formatting, guidance, constrained sampling and VQ decode while bounding the roughly 8.5B-parameter CPU run. Generation is limited to one prompt and one image; understanding, video, training, batches and fp16/bfloat16 parity are out of scope.
Memory. The published generator is float32 and occupies about 34 GB; its seven shards plus the vision tokenizer total about 35.7 GB on disk. CPU validation computes the reference in a subprocess and releases it before native inference.
License. The checkpoint, vision tokenizer and official code are Apache-2.0. imagehub does not distribute weights.