ORT — Ordinal-Biased Random Training

Class-conditional ImageNet-256 autoregressive generation checkpoints. Use the ORT code repository for model implementations, original AliTok/RAR baselines, training and inference. The code repository is public.

Available checkpoints

Model Epochs FID ↓ (paper) Download Recipe
ORT-E / AliTok-XL 400 1.18 Weights Config
ORT-E / AliTok-XL 300 1.26 Weights Config
ORT-L / AliTok-XL 300 1.34 Weights Config
ORT-L / AliTok-XL 400 1.31 Weights Config
AliTok-XL* baseline 300 1.42 Weights Config
AliTok-XL* baseline 400 1.35 Weights Config

FID values are paper-reported on ImageNet 256×256 (current manuscript, Tables 2 and 7). These rows correspond to the authors’ AliTok*-XL recipes. The released checkpoint-to-result mapping has not been independently verified by a new 50k-image evaluation. Exact per-row evaluation CFG and sampling seed remain unverified; config defaults and the example seed are not a claim to reproduce these FIDs.

Each file is 2.66 GB. Training seed: 42. SHA-256 manifest.

Download and generate

hf download donglixu/ORT ort-e-alitok-xl-400.bin configs/ort-e-alitok-xl-400.yaml --local-dir weights
python scripts/check_inference.py --config weights/configs/ort-e-alitok-xl-400.yaml --checkpoint weights/ort-e-alitok-xl-400.bin --tokenizer-weights weights/AliTok.pth --seed 2 --output outputs/example

Obtain the AliTok tokenizer from the official project. The generator predicts 273 tokens with vocabulary size 4096. The matching tokenizer converts them to a 256x256 RGB image. Sampling seed is independent of training seed. The checkpoint is a PyTorch state dictionary; load with torch.load(..., weights_only=True). This is a custom research model, not a Transformers AutoModel pipeline.

Validation and limitations

ORT-E 400 execution checks passed on GH200 with PyTorch 2.9.1: strict checkpoint loading, 273-token generation, tokenizer decoding and one synthetic optimizer update. No FID or full-dataset reproduction claim follows from these smoke tests. The config records the historical training recipe with portable paths and compilation disabled in the shared trainer. Inference CFG must be recorded explicitly in quantitative evaluations.

Weights contain no optimizer state, datasets, credentials or machine-specific metadata intentionally added by this release. Intended for research in image generation and reproducibility; generated images may contain artifacts and reflect dataset biases. Not validated for safety-critical decisions.

Attribution and terms

Based on AliTok, RAR, and LlamaGen. No blanket license is assigned to third-party code or tokenizer weights. AliTok-specific redistribution terms remain unresolved in the inspected upstream snapshot; consult upstream terms before reuse. Tokenizer weights are not re-hosted.

ORT-L 400’s portable recipe was recovered from its training log because the standalone config was stale. All tensors match the final step-250000 checkpoint. Other released files have source-to-Hub SHA-256 verification; their GPU inference and full FID evaluations have not been rerun for this release.

Reimplemented AliTok-XL baselines

* denotes models reimplemented and trained by us. These baselines use randomized pretraining with uniform token loss weights (alpha=beta=1), followed by raster-order post-training and testing. Use each downloaded config: model.generator.type: ort preserves the trained parameter layout. The official raster-order AliTok config is not interchangeable. Every tensor matches the final training checkpoint; parameter names and shapes match the verified ORT layout. Baseline GPU inference and FID have not been rerun for this release.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support