SceneWorks/pid-flux2

PiD — flux2 2kto4k decoder (SceneWorks redistribution)

A format-converted redistribution of NVIDIA's PiD (Pixel Diffusion) flux2 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 flux2 student serves the FLUX.2 VAE latent space (the packed 128-channel latent) — FLUX.2-dev, the FLUX.2-klein-9b variants, Lens (base/turbo), and Ideogram 4 (base/turbo).

⚠️ 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_flux2_2kto4k_v1pt5.safetensors	PiD v1.5 flux2 2kto4k 4-step student (461 tensors, bf16, ~2.8 GB). Improved color fidelity, no grid-corner artifacts, better anime/face detail. The default for the 2kto4k (up-to-4K) output tier, and it retires the v1.0 _2606 color-drift workaround — v1.5 fixes color fidelity at the source.
pid_flux2_2kto4k.safetensors	flux2 v1.0 2kto4k 4-step student (456 tensors, bf16, ~2.7 GB). Retained as the fallback for installs that predate v1.5.
pid_flux2_2k.safetensors	flux2 v1.0 res2k (2K-tuned) 4-step student. Sharper at 2048 px than the v1.5 2kto4k, so it serves the 2K output tier — NVIDIA ships no v1.5 res2k.
LICENSE	Full NVIDIA License text (verbatim from the original release).
NOTICE	NVIDIA attribution, provenance, and the exact conversion performed.

The engine picks v1.0 vs v1.5 by sniffing the weights (not the filename), so either 2kto4k file loads correctly, and an install with only the v1.0 file keeps working.

Provenance / conversion

Converted from NVIDIA's originals (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.

• v1.5 2kto4k: checkpoints/PiD_v1pt5_res2kto4k_sr4x_official_flux2_distill_4step/model_ema_bf16.pth
• v1.0 2kto4k: checkpoints/PiD_res2kto4k_sr4x_official_flux2_distill_4step_2606/model_ema_bf16.pth
• v1.0 res2k: checkpoints/PiD_res2k_sr4x_official_flux2_distill_4step/model_ema_bf16.pth

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