wenjieshu/lpgan

Latent Prior GANs

Efficient Decoder Refinement via Inherited Generative Structure

Decoder weights of LP-GANs. LP-GANs freeze the latent generator and image encoder, then adversarially refine the tokenizer decoder on generation-time latents with generation-aware paired supervision. Code, installation and evaluation: github.com/WenjieShu/Latent-Prior-GAN.

Each checkpoint stores the tokenizer state dict with the adapted decoder and is used with the corresponding frozen latent generator. ImageNet 256×256, 50,000 generated samples. gFID is reported under the OpenAI and official iMF evaluation protocols; lower is better. Values below are rounded as reported in the paper.

LP-GANs decoder checkpoints

Model	Evaluation config	OpenAI gFID ↓	Official iMF gFID ↓	Checkpoint
iMF-B/2	configs/imf_b.yaml	2.14	2.08	iMF-B-2-w-ps.pth
iMF-M/2	configs/imf_m.yaml	1.77	1.73	iMF-M-2-w-ps.pth
iMF-L/2	configs/imf_l.yaml	1.51	1.48	iMF-L-2-w-ps.pth
iMF-XL/2	configs/imf_xl.yaml	1.51	1.32	iMF-XL-2-w-ps.pth
MeanFlow-XL/2	configs/meanflow_xl.yaml	2.98	2.94	MeanFlow-XL-2-w-ps.pth
DMF-XL/2+	configs/dmf_xl.yaml	1.95	1.91	DMF-XL-2-plus-w-ps.pth
TiTok-L32	configs/titok_l32.yaml	1.98	1.90	TiTok-L32-w-ps.pth

The iMF-B/2 export is derived from the historical converted-PyTorch training checkpoint; its decoder tensors pass the public loader validation, while the source checkpoint is retained separately for provenance.

The continuous models use one-step sampling. iMF uses the paper's CFG evaluation settings; MeanFlow-XL/2 and DMF-XL/2+ use their native one-step samplers without additional sampling-time CFG. TiTok-L32 uses the frozen eight-step MaskGIT generator with native guidance scale 4.5. The reported values are historical paper results; the exported files preserve the original decoder tensors and precision and pass the public checkpoint loader.

Download

From the root of the code repository:

hf download wenjieshu/lpgan \
  --include "*.pth" \
  --local-dir checkpoints/lpgan

Usage

Download a decoder in Python:

from huggingface_hub import hf_hub_download

path = hf_hub_download("wenjieshu/lpgan", "iMF-XL-2-w-ps.pth")

For iMF, supply the corresponding official .pth prior and the SD-VAE, then run:

python eval_generation.py --config configs/imf_xl.yaml \
  --decoder_checkpoint /path/to/iMF-XL-2-w-ps.pth \
  --cfg_omega 8.0 --cfg_t_min 0.42 --cfg_t_max 0.62 \
  --num_samples 50000 --seed 42 --out_dir results/imf_xl_lpgan

MeanFlow-XL/2 uses the EMA VAE and its native pretrained generator. DMF-XL/2+ uses the MSE VAE and its native generator. TiTok-L32 uses the pretrained tokenizer and frozen MaskGIT generator from 1d-tokenizer:

python eval_generation.py --config configs/titok_l32.yaml \
  --tokenizer_ckpt_path /path/to/tokenizer_titok_l32.bin \
  --generator_ckpt_path /path/to/generator_titok_l32.bin \
  --decoder_checkpoint /path/to/TiTok-L32-w-ps.pth \
  --num_samples 50000 --seed 3407 --out_dir results/titok_lpgan

The continuous exports contain the adapted VAE decoder, frozen post-quantization convolution, and latent normalization buffers. The TiTok export contains the decoder, pixel quantizer, and pixel decoder. The original tokenizer or base VAE and frozen latent generator remain separate inputs. Optimizer and discriminator states are omitted. File checksums and metadata are in SHA256SUMS and checkpoints.json.

See the reproduction guide for pretrained inputs and evaluation details. The continuous models use sd-vae-ft-mse or sd-vae-ft-ema; see the upstream model cards and TiTok repository for their license terms.