SceneWorks/pid-sdxl

PiD — sdxl 2kto4k decoder (SceneWorks redistribution)

A format-converted redistribution of NVIDIA's PiD (Pixel Diffusion) sdxl 4-step distillation decoder, packaged for SceneWorks' native (no-PyTorch) PiD decoder (mlx-gen-pid on macOS/MLX, candle-gen-pid on Windows·Linux/CUDA).

PiD is an optional, per-generation replacement for the VAE decoder: it denoises directly in pixel space and decodes + 4× super-resolves in one 4-step pass. This sdxl student serves the SDXL VAE latent space (4-channel, the largest latent→pixel ratio) — SDXL base, RealVisXL (incl. RealVisXL Lightning), and Kolors, which share the SDXL VAE.

⚠️ License — non-commercial (research/evaluation) only

This checkpoint is under the NVIDIA License (HuggingFace tag NSCLv1); see LICENSE and NOTICE. Per §3.3, the Work and any derivative works may be used non-commercially — for research or evaluation purposes only (NVIDIA and its affiliates excepted).

This restriction flows to the output. Images decoded with PiD are for research/evaluation use only, distinct from the rest of the SceneWorks pipeline.

Contents

File	Description
pid_sdxl_2kto4k.safetensors	sdxl 2kto4k 4-step student backbone + sigma-aware LQ adapter (456 tensors, bf16, ~2.7 GB).
LICENSE	Full NVIDIA License text (verbatim from the original release).
NOTICE	NVIDIA attribution, provenance, and the exact conversion performed.

The PiD caption encoder (gemma-2-2b-it) and the SDXL VAE are provisioned separately by SceneWorks; this repo holds only the PiD student weights.

Provenance / conversion

Converted from NVIDIA's original checkpoints/PiD_res2kto4k_sr4x_official_sdxl_distill_4step/model_ema_bf16.pth (from nvidia/PiD) via a lossless key/format transform (net.-prefix strip + drop training-only net_ema.*/fake_score.*/discriminator.*; dtype preserved). No re-training.

• Project: https://research.nvidia.com/labs/sil/projects/pid/
• Code: https://github.com/nv-tlabs/pid
• Paper: arXiv:2605.23902