WPIT MAR-H — 800 epochs, EMA
EMA-only inference export of model_ema from checkpoint-799.pth (zero-based epoch 799 = 800 completed epochs). Original tensor dtypes are preserved; optimizer and non-EMA weights are not included.
Requires custom code from https://github.com/CHEN-YIZHU/WPIT_test and the matching WPIT tokenizer/decoder ckpt-last-90%WPIT.pth (not included). This file alone is not a complete image-generation pipeline and is not a standard Transformers model.
Instantiate the MAR-H architecture used by eval/mar_ray, then load with safetensors.torch.load_file and model.load_state_dict(state, strict=True). Architecture: image size 256, VAE stride 16, patch size 1, latent dimension 64, 1000 classes, buffer size 64, diffloss depth 12/width 1536, diffusion sampling steps 100. See training_args.json for the original configuration.
Verified evaluation
50,000 ImageNet class-conditional samples per setting. BF16, temperature 1, 256 MAR iterations, 100 diffusion steps, linear CFG, official ADM TensorFlow evaluator, IS split size 5000.
| Setting | gFID (lower better) | IS (higher better) |
|---|---|---|
| No CFG (scale 1) | 18.7584 | 107.0823 |
| CFG 3.2 | 16.5778 | 126.0364 |
UniFlow paper settings differ in generator size, training schedule and resolution; these are not controlled tokenizer-only comparisons. Upstream code, base model and dataset terms continue to apply; no additional license grant is asserted here.
All 604 EMA tensors passed finiteness and exact save/load equality checks. SHA256SUMS records the exported file hash.