prologue

Prologue — Autoregressive Visual Generation Needs a Prologue

Paper: arXiv 2605.06137 · Code: github.com/Zyriix/prologue · Demo: 🤗 Space

Bowen Zheng · Weijian Luo · Guang Yang · Colin Zhang · Tianyang Hu

Image = [prologue tokens] + [visual tokens]. Prologue tokens are a small set of latent tokens prepended to the visual sequence and trained only with the AR cross-entropy loss. Visual tokens stay dedicated to reconstruction. The reconstruction–generation gap closes for free, and prologue tokens spontaneously develop semantic structure under pure CE gradients.

The figure above is generated with fixed prologue tokens, varying visual tokens only (Prologue-L–XL). Class identity and global layout stay; texture varies.

---

Headline numbers — ImageNet 256×256

All numbers are reproducible end-to-end with bash eval.sh in the GitHub repo. gFID / IS follow the ADM evaluation protocol (50k samples vs VIRTUAL_imagenet256_labeled.npz).

---

What's in this repo

This Hugging Face Hub repository hosts all released weights for the paper — 6 tokenizers and 9 AR models (~63 GB total). The matching code, training scripts, and full documentation live on GitHub.

Tokenizers (6 directories)

AR models (9 directories)

All checkpoints are safetensors following the 🤗 Accelerate convention (model.safetensors, model_1.safetensors, …). After download the layout matches the relative paths expected by eval.sh / app.py in the code repo — no mv step required.

---

Download

pip install -U "huggingface_hub[cli]"
export HF_XET_HIGH_PERFORMANCE=1   # parallel Xet transfer

# everything (~63 GB)
hf download Zyriix/prologue --local-dir ckpts

# or just the headline model used by the demo (LXL, 9.9 GB + 6.7 GB tokenizer)
hf download Zyriix/prologue \
    --include "ar-prologue-l-xl/*" \
    --include "prologue-l-tokenizer/*" \
    --local-dir ckpts

See the GitHub README for per-model commands and an inference-only slim layout (drops the ~50 % of bytes used for resuming training).

---

How to use

git clone https://github.com/Zyriix/prologue.git && cd prologue
bash setup_env.sh && conda activate prologue

# unpack the released ckpts (see above)
hf download Zyriix/prologue \
    --include "ar-prologue-l-xl/*" \
    --include "prologue-l-tokenizer/*" \
    --local-dir ckpts

# (a) full headline-table reproduction
bash eval.sh

# (b) interactive Gradio demo: fix prologue, resample visual
python app.py

Programmatic loading inside Python:

from huggingface_hub import snapshot_download
ckpt_dir = snapshot_download(
    repo_id="Zyriix/prologue",
    allow_patterns=["ar-prologue-l-xl/*", "prologue-l-tokenizer/*"],
    local_dir="ckpts",
    max_workers=8,
)
# Then call into prologue/ as a library (load_models / sample_tokens):
#   from sample_vis import load_models, sample_tokens
#   See app.py for a full minimal example.

---

Model details

• Architecture. Causal Transformer AR over a prologue prefix (z_len = 16, separate codebook) and a 2D visual sequence (256 tokens, x_len = 16² × 1, codebook = 4096 or 16 384). Conditioning is class-label one-hot injected as a BOS embedding (LlamaGen style). RoPE positions, RMSNorm, GeGLU FFN.
• Tokenizer. ViT-style encoder/decoder with simvq codebook (Yu et al.). Prologue variants share the encoder; Prologue-Post freezes a 2D visual tokenizer and adds a second prologue path on top.
• Training stage 1 — tokenizer + joint AR head. L1 + LPIPS (or ConvNeXt-logit for L variants) + PatchGAN. The AR head is trained with cross-entropy on [prologue ; visual]; STE through the prologue codebook flows gradients into the encoder. Base: 150 epochs / Large: 200 epochs, both at batch size 256.
• Training stage 2 — pretokenize. One pass over ImageNet, cache token indices to sharded .npz.
• Training stage 3 — large AR. Cross-entropy only, on the cached tokens. AR-Base: 400 epochs at batch 512 · AR-Large: 800 epochs at batch 2048, aligned with AliTok. Supports separate temperatures and separate CFG schedules for prologue / visual.
• Numerics. BF16 mixed precision, torch.compile, flash-attn 2.8 (source-built against torch 2.9.1 + cu128).
• Compute. Paper numbers were run on 8× H100 80 GB; the released training and eval scripts auto-scale to any (1× or 8×) configuration via 🤗 Accelerate.

---

Intended use, limitations, and risks

• Intended use. Research on discrete image tokenization, AR image generation, and the role of semantic prefixes in token sequences. The released models are class-conditional on the 1 000 ImageNet classes only — there is no text conditioning.
• Out-of-distribution behaviour. Samples for classes that are visually similar to ImageNet categories are best; the model has no notion of arbitrary text prompts, layout, or style.
• Failure modes. Like all AR image models, the model occasionally produces anatomical / object artifacts, particularly under aggressive (high-CFG) sampling. Prologue prefixes constrain global layout but not fine-grained correctness.
• Bias. ImageNet-1k itself contains known social and cultural biases; downstream generation inherits them. Do not deploy these weights for any application affecting individuals' lives, in any high-stakes setting, or as input to face-related pipelines.
• License. Apache-2.0 for the weights and the bulk of the code. See NOTICE for the CC BY-NC-SA 4.0 carve-out on four NVIDIA StyleGAN3-derived files.

---

Citation

@article{zheng2026prologue,
  title   = {Autoregressive Visual Generation Needs a Prologue},
  author  = {Zheng, Bowen and Luo, Weijian and Yang, Guang and Zhang, Colin and Hu, Tianyang},
  journal = {arXiv preprint arXiv:2605.06137},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.06137}
}

Acknowledgements

Inspired by (chronological) LPIPS (2018) · vector-quantize-pytorch (2020) · VQGAN / taming-transformers (2020) · guided-diffusion (2021) · VAR (2024.04) · LlamaGen (2024.06) · TiTok (2024.06) · Open-MAGVIT2 (2024.09) · ImageFolder (2024.10) · AliTok (2025.06).