Nucleus-Image¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
nucleus-image |
NucleusAI/Nucleus-Image | 5e963db |
Apache-2.0 | Verified |
Catalog: nucleus-image · Models: nucleus-image · Task: t2i
Architecture. Nucleus-Image combines a Qwen3-VL text encoder, a 32-block single-stream diffusion transformer, expert-choice SwiGLU mixture-of-experts layers, grouped-query attention, three-axis RoPE, and the Qwen-Image VAE. Published weights are bfloat16.
Settings. The backend supports steps, guidance, negative_prompt, width, height,
seed, latents, and max_sequence_length. Dimensions must be multiples of 16.
Validation. On CPU in float32, the same seed and settings give the same 1024×1024 image as
diffusers' NucleusMoEImagePipeline with generator=torch.Generator().manual_seed(seed)
(validation). The expert layers follow the published use_grouped_mm setting on
CPU and CUDA. MPS and CUDA are not validated.