Research map / Architectures and efficiency / Quantization
RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization
Kaicheng Yang, Xun Zhang, Haotong Qin, Yucheng Lin, Kaisen Yang, Xianglong Yan, Yulun Zhang
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 |
|---|---|---|---|
| add low-rank branch (LRB) to BitNetv2 (ablation) | BitNetv2 W1.58A4 · ImageNet 256x256 W1.58A4 (cfg 1.5) | FID: 41.59 -> 29.59 | Improved |
| add Hadamard transform to all linear layers on top of LRB (ablation) | BitNetv2 + LRB W1.58A4 · ImageNet 256x256 W1.58A4 (cfg 1.5) | FID: 29.59 -> 20.82 | Improved |
| add uniform per-token Gauss quantizer (ablation) | Baseline W1.58A4 · ImageNet 256x256 W1.58A4 (cfg 1.5) | FID: 20.82 -> 17.97 | Improved |
| add non-uniform per-token Gauss quantizer (ablation) | Baseline W1.58A4 · ImageNet 256x256 W1.58A4 (cfg 1.5) | FID: 20.82 -> 20.33 | No clear effect |
| add activation-only mixed precision (AMPN) at W1.58A3 (ablation) | RobuQ w/o AMP · ImageNet 256x256 W1.58A3 (cfg 1.5) | FID: 24.44 -> 21.40 | Improved |
| add activation-only mixed precision (AMPN) at W1.58A2 (ablation) | RobuQ w/o AMP · ImageNet 256x256 W1.58A2 (cfg 1.5) | FID: 43.31 -> 30.30 | Improved |
| raise QAT steps for metric collection from 1 to 1,000 (ablation) | AMPN W1.58A2 · ImageNet 256x256 W1.58A2 (cfg 1.5) | FID: 41.06 -> 30.30 | Improved |
| raise QAT steps for metric collection from 1,000 to 1,500 (ablation) | AMPN W1.58A2 · ImageNet 256x256 W1.58A2 (cfg 1.5) | FID: 30.30 -> 30.56 | No clear effect |
| apply RobustQuantizer W1.58A4 instead of baseline on FFHQ | Baseline W1.58A4 · FFHQ 256x256 (unconditional) | FID: 34.32 -> 25.62 | Improved |
| apply RobuQ W1.58A2 instead of baseline on ImageNet cfg 4.0 | Baseline W1.58A2 · ImageNet 256x256 (cfg 4.0) | FID: 75.02 -> 11.06 | Improved |
| W4A4 PTQ with RobuQ on FLUX-Schnell | FLUX-Schnell · MJHQ (4 steps) | FID: 18.40 -> 18.57 | No clear effect |
| W4A4 PTQ with RobuQ on alpha-Flow at NFE=1 | alpha-Flow-XL/2 · ImageNet-1K (NFE=1) | FID: 2.58 -> 2.59 | No clear effect |
Closest papers by meaning
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.93)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.92)
- DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.90)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.89)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 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.
