Research map / Image generation / Autoregressive and latent generation

Balancing Image Compression and Generation with Bootstrapped Tokenization

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Generative latent space tooling · 0 citations (OpenAlex)

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 SelfBootTok, a 1D image tokenizer that splits tokens into global and local groups and self-bootstraps local details from global tokens, reaching gFID 1.56 on ImageNet-256 with 64 tokens.

Problem

Standard image tokenizers mix all granularities within each token, so redundancy persists between tokens and generators must produce all visual detail at once, which limits generation efficiency and tokenizer scalability.

Approach

Decompose tokens into global (semantic) and local (1D and 2D) groups with separate codebooks; local tokens are predicted from global tokens by an MLP and a causal transformer aligner, trained with pretrained encoder alignment and an entropy-regularized optimal transport loss mapping 2D VAE latents to the 1D sequence. The generator only produces global tokens, and local aligners can be scaled witho

Key results

• SelfBootTok with 64 tokens achieves gFID 1.56 (with CFG) on ImageNet-256, the best among 1D tokenizers in the paper's comparison table.
• Reconstruction rFID of 0.66 with 64 tokens, versus 0.88 for SoftVQ with the same token count.
• Scaling the local aligner reduces total training computation by about 40% (4.9B to 2.95B params; 60K to 36.6K GFLOPs) and training time by about 54% (28 to 13 days).

Surprising findings

• Concatenation fusion improves reconstruction but degrades generation quality, indicating a task-dependent trade-off.

Limitations

• Gains over the strongest 2D baselines are reported mainly at 64 tokens; VA-VAE with 256 tokens still reaches a lower gFID of 1.35.
• Generator is evaluated only on class-conditional ImageNet-256 generation.

Experimental setup

Datasets: ImageNet-1K

Benchmarks: ImageNet-256

Metrics: rFID gFID (with CFG) gFID (without CFG) PSNR SSIM GFLOPs

Baselines: TiTok SoftVQ FlexTok GigaTok MAETok HieraTok VA-VAE MAR-H LightningDiT

Builds on: TiTok SoftVQ VA-VAE SiT REPA

Closest papers by meaning

• When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
• An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.87)
• Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.86)
• FlexTok: Resampling Images into 1D Token Sequences of Flexible Length (ICML 2025 · similarity 0.86)
• Highly Compressed Tokenizer Can Generate Without Training (ICML 2025 · similarity 0.86)
• Language-Guided Image Tokenization for Generation (CVPR 2025 · similarity 0.86)