Ana içeriğe geç

PixelDiT

Model Source Revision Weight license Parity (CPU, fp32)
pixeldit nvidia/PixelDiT-1300M-1024px 7c63b99 NVIDIA License (NSCLv1; non-commercial research or evaluation) Verified

Catalog: pixeldit · Models: pixeldit · Task: t2i

Architecture. PixelDiT generates RGB pixels directly, without an autoencoder. Fourteen patch-level MM-DiT blocks jointly attend to image patches and Gemma-2 text tokens for global semantics. Two pixel-level transformer blocks then compact each 16×16 patch into one attention token and expand it back to pixel tokens, with per-pixel AdaLN modulation from the patch stream.

Usage.

python -m imagehub.pretrained pixeldit checkpoints/pixeldit
from imagehub import pipeline

with pipeline("t2i", backend="pixeldit", model="checkpoints/pixeldit") as pipe:
    result = pipe("a blue ceramic cat on a wooden table", seed=0)
result.images[0].save("cat.png")

Settings. Defaults are 50 steps, guidance 2.75, seed 0 and 1024×1024. width and height must be divisible by 16. negative_prompt, interval_guidance and use_chi_prompt are also supported. Sampling uses the official second-order flow DPM-Solver++ with flow shift 4.

Validation. The 391 published PixelDiT tensors and both Gemma-2 shards have strict layout tests. Random-weight transformer, flow-solver and tiny end-to-end comparisons against the pinned official implementation are bit-exact. The opt-in real-checkpoint suite covers the tokenizer, Gemma-2 output, one denoiser step, and separate seed-0 and seed-42 runs through the public pipeline; the end-to-end cases use two steps at 16×16 to keep CPU execution short while exercising the full pixel-space path. Checkpoint tensors are copied from pickle storage before assignment to preserve the parameter layout and CPU summation order used by the published loader.

Memory and limits. The transformer has about 1.3B parameters and Gemma-2 about 2.6B. Training, class-conditional checkpoints, FlowEdit, batches, caller-provided noise and reduced-precision parity are out of scope.

License. The weights and official implementation use the NVIDIA NSCLv1 license and are limited to non-commercial research or evaluation. imagehub does not distribute weights.