Research map / Vision-language and multimodal / Human and embodied modeling
SPAE: Semantic Pyramid AutoEncoder for Multimodal Generation with Frozen LLMs
Lijun Yu, Yong Cheng, Zhiruo Wang, Vivek Kumar, Wolfgang Macherey, Yanping Huang, David A Ross, Irfan Essa, Yonatan Bisk, Ming-Hsuan Yang, Kevin Patrick Murphy, Alexander G Hauptmann and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPAE lexical tokens instead of LQAE tokens | GPT 3.5 few-shot classification · mini-ImageNet 2-way classification | average accuracy: 53.97 -> 69.51 | Improved |
| Use SPAE lexical tokens instead of LQAE tokens | GPT 3.5 few-shot classification · mini-ImageNet 5-way classification | average accuracy: 29.04 -> 53.91 | Improved |
| Use SPAE with PaLM 2 rather than Frozen | few-shot VQA · Real-Fast-VQA, 1 inner shot | VQA performance: 7.8 -> 14.3 | Improved |
| Replace the learned codebook with a frozen language codebook (ablation) | SPAE tokenizer · ImageNet ILSVRC2012 128×128 | FID: 5.48 -> 7.44 | Hurt |
| Add semantic loss to the frozen codebook (ablation) | SPAE tokenizer · ImageNet ILSVRC2012 128×128 | FID: 7.44 -> 5.17 | Improved |
| Use 2-layer RQ with the frozen codebook (ablation) | SPAE tokenizer · ImageNet ILSVRC2012 128×128 | FID: 5.17 -> 11.94 | Hurt |
| Use 2-layer SAQ instead of 2-layer RQ (ablation) | SPAE tokenizer · ImageNet ILSVRC2012 128×128 | FID: 11.94 -> 12.30 | Hurt |
| Add six-layer pyramid SAQ (ablation) | SPAE tokenizer · mini-ImageNet 5-way 1-shot | classification accuracy: 46.2 -> 65.1 | Improved |
| Remove per-layer semantic thresholds (ablation) | SPAE · ImageNet ILSVRC2012 128×128 | FID: 4.41 -> 4.33 | Improved |
| Remove dynamic semantic loss weight (ablation) | SPAE · ImageNet ILSVRC2012 128×128 | FID: 4.41 -> 9.00 | Hurt |
| Remove perceptual loss (ablation) | SPAE · ImageNet ILSVRC2012 128×128 | FID: 4.41 -> 40.47 | Hurt |
| Remove perceptual loss (ablation) | SPAE · mini-ImageNet 5-way 1-shot | classification accuracy: 65.1 -> 69.5 | Improved |
Later papers that compared with this method
1 later accepted paper reports SPAE as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Beyond Text: Frozen Large Language Models in Visual Signal Comprehension (CVPR 2024 · better on 10 of 10 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LMFusion: Adapting Pretrained Language Models for Multimodal Generation (NeurIPS 2025 · similarity 0.89)
- Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs (NeurIPS 2026 · similarity 0.88)
- MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer (ICLR 2026 · similarity 0.88)
- SynerGen-VL: Towards Synergistic Image Understanding and Generation with Vision Experts and Token Folding (CVPR 2025 · similarity 0.88)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 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.
