Ana içeriğe geç

Lumina-DiMOO

Model Source Revision Weight license Parity (CPU, fp32)
lumina-dimoo Alpha-VLLM/Lumina-DiMOO f014549 Apache-2.0 Verified

Catalog: lumina-dimoo · Model: lumina-dimoo · Task: t2i

Architecture. Lumina-DiMOO is a fully discrete diffusion model. Its 32-layer, width-4096 LLaDA transformer uses bidirectional attention, 32 heads, rotary position embeddings, RMS norms and gated SiLU feed-forwards. Text and 8,192 image tokens share one 134,548-entry vocabulary. Generation appends a 16x-downsampled rectangular grid of mask tokens to the formatted prompt. At each step the transformer predicts every still-masked image token, Gumbel-max samples it, and a cosine schedule remasks the least-confident samples. The final ids are decoded by the published aMUSEd VQ model.

Usage.

python -m imagehub.pretrained lumina-dimoo checkpoints/lumina-dimoo
from imagehub import pipeline

with pipeline("t2i", backend="lumina-dimoo", model="checkpoints/lumina-dimoo") as pipe:
    result = pipe("a red cube on a white table", seed=42)
    result.images[0].save("cube.png")

Settings. Defaults follow the published command-line sampler: 64 steps, guidance 4.0, temperature 1.0 and 1024x1024 output. width and height may be non-square positive multiples of 16. Supported options are seed, steps, guidance, temperature, width and height. Temperature 0 uses greedy token selection. Guidance 0 skips the unconditional forward pass. Unknown options are rejected before loading.

Memory and validation. The four language-model shards contain BF16 weights (about 15 GB); the VQ decoder is float32 (about 0.6 GB). CPU float32 is the parity target. Header checks cover all 291 LLaDA and 277 VQ tensors. Random-weight LLaDA logits and tiny masked-token trajectories are exact; the real-checkpoint tests use 32x32 images and two steps to keep their CPU run short.

Out of scope. Image editing, inpainting, understanding, batching, the experimental max-logit cache, training, and reduced-precision parity are not supported.

License. The code and weights are Apache-2.0. imagehub does not distribute weights.