Show-o¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
show-o-512 |
showlab/show-o-512x512 | dbe757d |
Apache-2.0 | Verified |
Catalog: show-o ·
Models: show-o-512 · Task: t2i
Architecture. Show-o extends the 1.3B Phi-1.5 decoder to a shared vocabulary containing text, prompt-control and 8,192 image tokens. Text uses causal attention, while the 1,024 image-token region uses bidirectional attention. Image tokens are predicted in parallel over 18 cosine-remasking rounds and decoded by a lookup-free MAGVIT-v2 tokenizer with 16× spatial compression.
Usage.
from imagehub import pipeline
with pipeline("t2i", backend="show-o-512", model="checkpoints/show-o-512") as pipe:
result = pipe("a watercolor painting of a lighthouse", seed=42)
result.images[0].save("lighthouse.png")
Settings. The checkpoint generates 512×512 images. Defaults are 18 steps, guidance 10 and
temperature 1. The supported options are seed, steps, guidance, temperature, width and
height; width and height must remain 512. Unrecognized settings raise an error before weights load.
Memory and validation. The float32 unified transformer is about 5.8 GB and MAGVIT-v2 about 0.38 GB. Offline tests cover every published tensor, CodeGen tokenization, Phi attention and MLP, MAGVIT-v2 decoding, omni-attention prompting, discrete denoising, and short two-step end-to-end runs at seeds 0 and 42. The shortened run changes only the step count from the 18-step default.
Out of scope. Multimodal understanding, image input, inpainting, extrapolation, CLIP-ViT variants, 256-pixel checkpoints, batches and reduced-precision parity are not supported.
License. The model card declares Apache-2.0. imagehub does not distribute weights.