Ovis-Image¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
ovis-image-7b |
ATH-MaaS/Ovis-Image-7B | 41be1c5 |
Apache-2.0 | Verified |
Catalog: ovis-image · Models:
ovis-image-7b · Task: t2i
Ovis-Image uses a Qwen3 text encoder and a flow-matching joint transformer, followed by a
16-channel AutoencoderKL. It accepts text prompts and defaults to 50 steps, guidance 5.0, and
1024×1024 pixels. seed, steps, guidance, width, height, negative_prompt,
num_images, and max_sequence_length are supported.
The published transformer and text encoder are float32; the VAE is bfloat16. CPU float32 parity and end-to-end validation are covered; the real-checkpoint run took 12 minutes on an M2 Max (CPU) (validation).