Beyond Selection: Token Parameterization for Extreme Visual Token Compression
Braco compresses visual tokens 23–144× while preserving VLM accuracy and cutting prefill FLOPs.
Researchers introduce Braco, a lightweight four-step coder for extreme visual-token compression in vision-language models. It combines transform-basis truncation, input-independent basis-coordinate embeddings, budget-dependent orthogonal re-parameterization, and learned spatial residual tokens. Braco forms a favorable accuracy-efficiency frontier at 23×–64× compression and remains competitive at 144×, reaching 95.2% accuracy while reducing prefill FLOPs by 84.2%–86.7%. Against prior methods it matches or improves accuracy with up to about 36% end-to-end speedup and 16.6×/78.8× lower compressor latency/FLOPs.
- Braco separates basis truncation from coordinate organization for visual tokens.
- It reaches 95.2% accuracy while cutting prefill FLOPs by 84.2%–86.7%.
- Competitive from 23× through 144× compression versus prior methods.
- Up to about 36% end-to-end speedup and much lower compressor cost.
Full article157 words · extracted from huggingface.co · click to collapse
Visual-token compression is effective for improving the efficiency of vision-language models, but under extreme compression budgets, token pruning can break visual grounding while learned resamplers increase parameter count, attention cost, and training complexity. We revisit compression through a token parameterization lens, separating (i) basis transformation and structured truncation (retained subspace/compressibility) from (ii) coordinate organization (optimization and cross-modal alignment). This view yields two coupled objectives, compressibility and learnability, which we formalize as unified functionals. Guided by these objectives, we design Braco, a lightweight four-step coder that combines transform-basis truncation, input-independent basis-coordinate embeddings, budget-dependent orthogonal re-parameterization, and learned spatial residual tokens from lightweight pooling. Experiments show that Braco forms the favorable empirical accuracy-efficiency frontier under 23times--64times compression and remains competitive at 144times, reaching 95.2% accuracy while reducing prefill FLOPs by 84.2%--86.7% relative to the uncompressed upper bound. Against prior methods, Braco matches or improves accuracy while achieving up to approximately 36% end-to-end speedup and using 16.6times/78.8times lower compressor latency/FLOPs.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.35232