Research map / Image generation / Fast diffusion sampling
BiGain: Unified Token Compression for Joint Generation and Classification
Jiacheng Liu, Shengkun Tang, Jiacheng Cui, Dongkuan Xu, Zhiqiang Shen
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 Laplacian-gated merging instead of ToMe | Stable Diffusion 2.0 diffusion classifier · Pets (10% FLOPs reduction) | Acc@1 (%): 72.96 -> 78.38 | Improved |
| use Laplacian-gated merging instead of ToMe at 70% merge | Stable Diffusion 2.0 · ImageNet-1K (2,000 images) | Acc@1 (%): 37.35 -> 44.50 | Improved |
| use Laplacian-gated merging instead of ToMe at 70% merge | DiT-XL/2 · ImageNet-100 | Acc@1 (%): 53.88 -> 61.76 | Improved |
| use Laplacian-gated merging instead of ToMe at 70% merge | Stable Diffusion 2.0 · Pets generation | FID: 38.35 -> 37.73 | Improved |
| use IE-KVD instead of ToDo at 4x downsampling | Stable Diffusion 2.0 · ImageNet-100 | Acc@1 (%): 57.62 -> 61.72 | Improved |
| use IE-KVD instead of ToDo at 3x downsampling | DiT-XL/2 · ImageNet-100 | Acc@1 (%): 8.46 -> 61.58 | Improved |
| merge cross-attention tokens in addition to self-attention (ablation) | BiGain_TM · Pets (70% merge) | Acc@1 (%): 74.63 -> 73.89 | Hurt |
| merge MLP tokens in addition to self-attention (ablation) | BiGain_TM · Pets (70% merge) | Acc@1 (%): 74.63 -> 68.27 | Hurt |
| merge cross-attention and MLP tokens in addition to self-attention (ablation) | BiGain_TM · Pets (70% merge) | Acc@1 (%): 74.63 -> 68.25 | Hurt |
| score tokens with Laplacian L1 instead of global mean deviation (ablation) | BiGain_TM · Pets (70% merge) | Acc@1 (%): 72.96 -> 74.63 | Improved |
| score tokens with Laplacian L1 instead of DFT spectral centroid (ablation) | BiGain_TM · Pets (70% merge) | Acc@1 (%): 73.75 -> 74.63 | Improved |
| keep only low-frequency KV tokens instead of nearest-neighbor downsampling (ablation) | ToDo-style KV downsampling · ImageNet-100 | Acc@1 (%): 72.30 -> 45.58 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MosaicDiff — MosaicDiff: Training-free Structural Pruning for Diffusion Model Acceleration Reflecting Pretraining Dynamics (ICCV 2025) · better on 1 of 1 shared result
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
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.90)
- Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation (NeurIPS 2026 · similarity 0.90)
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.89)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.89)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.89)
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.
