hzeroyuke/vae-posttraining

VAE checkpoints for ImageNet LlamaGen GPT-XL

This repository contains the two VQ-16 class-conditional ImageNet tokenizers used by the paired LlamaGen GPT-XL experiment:

• base_vae_vq_ds16_c2i.pt: official FoundationVision/LlamaGen VQ-16 c2i checkpoint.
• reward_vae_vq_ds16_c2i_seed3101_step1000.pt: reward-tuned VQ-16 c2i checkpoint, exported from the seed 3101, step 1000 tokenizer run.

Both checkpoints use a 16,384-entry codebook and produce 24 x 24 = 576 tokens for a 384px ImageNet input. They are intended for the ImageNet c2i training pipeline in https://github.com/hzeroyuke/vae-posttraining.

SHA256

File	SHA256
base_vae_vq_ds16_c2i.pt	109aa8afb2cf3761eec23cdc8644154cb498f5ab7eef2a35264d25e5e0499f7d
reward_vae_vq_ds16_c2i_seed3101_step1000.pt	55c06b0ab914f9f40b24eba07033f7c452eecb6233562baf27e69121282434ea

The checkpoints are distributed for research use. Follow the licenses and terms of LlamaGen, ImageNet, PyTorch, and any downstream evaluation dependencies.