Research map / Image generation / Autoregressive and latent generation
BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BiGR, a class-conditional masked generative model over binary latent codes with a Bernoulli-diffusion binary transcoder and entropy-ordered sampling, which achieves competitive ImageNet generation and strong linear-probe representations.
Problem
Conditional generative models produce good images but their learned representations are weak for discriminative tasks; the paper asks whether one conditional generator can perform both generation and linear-probe classification well.
Approach
Model compact binary latent codes (from a binary autoencoder) with a bidirectional Llama-style masked transformer, predict bits via a Bernoulli denoising binary transcoder trained with weighted BCE on XOR residuals, sample by unmasking tokens in order of binary-entropy confidence, and use average-pooled intermediate features as representations.
Key results
- BiGR-XXL-d32 reaches FID 2.36 on ImageNet 256x256 generation, with BiGR-L-d24 at 2.71 (336M params).
- BiGR-XL-d32 achieves 69.8% linear-probe ACC1 on ImageNet, versus 40.5% for LlamaGen-L and 60.0% for MAR-H among conditional generative models.
- In the uniformity comparison, the masked binary model S3 reaches FID 3.17 and ACC1 64.3, versus LlamaGen's FID 3.81 and ACC1 40.5, while sampling in 0.69 s versus 0.13 s for LlamaGen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BiGR instead of LlamaGen | conditional ImageNet generation and representations · ImageNet-1K 256×256 | FID: 3.81 -> 3.17 | Helped |
| use BiGR instead of LlamaGen | conditional ImageNet generation and representations · ImageNet-1K 256×256 | ACC1: 40.5 -> 64.3 | Helped |
| replace categorical loss with binary loss under masked modeling (ablation) | BiGR image generation · ImageNet-1K 256×256 | FID: 3.85 -> 3.17 | Helped |
| replace masked modeling with autoregressive modeling under binary loss (ablation) | BiGR image generation · ImageNet-1K 256×256 | FID: 7.50 -> 3.17 | Helped |
| replace Bernoulli denoising with direct BCE (ablation) | BiGR-L-d16 · ImageNet-1K 256×256 | FID: 3.17 -> 5.84 | Hurt |
| predict the initial clean latent z^0 instead of the XOR residual (ablation) | BiGR-L-d16 · ImageNet-1K 256×256 | FID: 3.17 -> 4.39 | Hurt |
| use raster-scan unmasking order instead of entropy order (ablation) | BiGR-L-d16 · ImageNet-1K 256×256 | FID: 3.17 -> 4.51 | Hurt |
| use random unmasking order instead of entropy order (ablation) | BiGR-L-d16 · ImageNet-1K 256×256 | FID: 3.17 -> 7.12 | Hurt |
| use deterministic thresholding instead of Bernoulli sampling (ablation) | BiGR-L-d16 · ImageNet-1K 256×256 | FID: 3.17 -> 3.19 | Hurt |
| replace class-condition tokens with a single unconditional token (ablation) | BiGR-L-d20 · ImageNet-1K | linear-probe accuracy | Helped |
| reduce diffusion timesteps to 10 (ablation) | BiGR-XXL · ImageNet-1K generation | FID: 2.73 | No clear effect |
| use 10 sampling iterations instead of the usual 20–30 (ablation) | BiGR generation · ImageNet-1K generation | generation quality | Hurt |
Surprising findings
- Unconditional training yields better linear-probe representations than the class-conditional counterpart, suggesting conditioning hurts discriminability.
- Direct BCE without Bernoulli denoising produces overly smooth images and worse generation quality.
Limitations
- Sampling strategy has many hyperparameters, so the reported settings may not be optimal.
- Fixed training sequence length means higher-resolution generation requires retraining.
Experimental setup
Datasets: ImageNet-1K LAION-400M JourneyDB
Benchmarks: ImageNet-1K 256x256 generation (FID-50k, IS, sFID, Precision, Recall) ImageNet-1K linear probe (ACC1, ACC5)
Metrics: FID IS sFID Precision Recall ACC1 ACC5 inference time
Baselines: LlamaGen DiT SiT MAR VAR MaskGIT MAE MAGE iBOT ViT-VQGAN
Builds on: LlamaGen MaskGIT Binary autoencoder (B-AE) Bernoulli diffusion Llama MAR
Closest papers by meaning
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
- Don't Look into the Dark: Latent Codes for Pluralistic Image Inpainting (CVPR 2024 · similarity 0.87)
- Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.87)
- Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.