PixelDiT / PixelDiT2
Copyright (c) NVIDIA Corporation

This repository is released under the license in LICENSE. It contains code
derived from the third-party projects listed below; their terms apply to the
corresponding files.

--------------------------------------------------------------------------------
PixNerd — https://github.com/MCG-NJU/PixNerd

  c2i/main.py
  c2i/src/data.py
  c2i/src/diffusion.py
  c2i/src/lightning.py
  c2i/src/utils.py
  pixdit_core/modules.py
  pixdit_core/pixeldit_c2i.py

--------------------------------------------------------------------------------
JiT — https://github.com/LTH14/JiT
MIT License, Copyright (c) 2025 Tianhong Li

  The PixelDiT2 backbone conventions (in-context class tokens, x-prediction with
  a velocity-space loss, the clipped velocity conversion) and the ODE samplers
  derive from this project.

  pixdit_core/pixeldit2_c2i.py
  c2i/src/grounded_diffusion.py

--------------------------------------------------------------------------------
guided-diffusion (ADM) — https://github.com/openai/guided-diffusion
MIT License, Copyright (c) 2021 OpenAI

  center_crop_arr in c2i/src/data.py.
  The FID protocol follows the ADM evaluation suite.

--------------------------------------------------------------------------------
REPA / REPA-E — https://github.com/sihyun-yu/REPA,
                https://github.com/End2End-Diffusion/REPA-E

  The representation alignment loss and the ImageNet preprocessing pipeline.

--------------------------------------------------------------------------------
SANA — https://github.com/NVlabs/SANA
Apache License 2.0

  The t2i/ subtree; per-file SPDX headers apply.

--------------------------------------------------------------------------------
Pretrained encoder weights

PixelDiT2 loads frozen vision encoders at runtime through `timm`; their weights
are not redistributed here and are governed by their own licenses.

  DINOv3 (grounding encoder) — Meta Platforms, DINOv3 License
    timm/vit_small_patch16_dinov3_qkvb.lvd1689m
    timm/vit_base_patch16_dinov3_qkvb.lvd1689m
    timm/vit_large_patch16_dinov3_qkvb.lvd1689m

  DINOv2 (REPA target, training only) — Meta Platforms, Apache License 2.0
    timm/vit_base_patch14_dinov2.lvd142m

The released PixelDiT2 checkpoints do not contain encoder weights. The frozen
encoder weights are loaded from timm at runtime.

--------------------------------------------------------------------------------
FID reference statistics

The ImageNet-256 and ImageNet-512 reference statistics used to reproduce the
reported FID values are the files tracked in the JiT repository
(https://github.com/LTH14/JiT, fid_stats/jit_in256_stats.npz and
fid_stats/jit_in512_stats.npz). They are not redistributed here.
