Research map / Image generation / Autoregressive and latent generation
Cubic Discrete Diffusion: Discrete Visual Generation on High-Dimensional Representation Tokens
Yuqing Wang, Chuofan Ma, Zhijie Lin, Yao Teng, Lijun Yu, Shuai Wang, Jiaming Han, Jiashi Feng, Yi Jiang, Xihui Liu
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 |
|---|---|---|---|
| mask individual elements across the tensor instead of whole spatial positions (ablation) | CubiD-L · ImageNet 256x256 | gFID: 22.22 -> 5.33 | Improved |
| mask individual elements instead of whole dimensions across positions (ablation) | CubiD-L · ImageNet 256x256 | gFID: 120.03 -> 5.33 | Improved |
| use a learned mask token instead of a random codebook token (ablation) | CubiD-L · ImageNet 256x256 | gFID: 56.38 -> 5.33 | Improved |
| use a learned mask token instead of a fixed zero token (ablation) | CubiD-L · ImageNet 256x256 | gFID: 5.56 -> 5.33 | Improved |
| use mask ratio std sigma=0.10 rather than 0.05 (ablation) | CubiD-L · ImageNet 256x256 | gFID: 7.65 -> 5.33 | Improved |
| use mask ratio std sigma=0.10 rather than 0.15 (ablation) | CubiD-L · ImageNet 256x256 | gFID: 5.81 -> 5.33 | Improved |
| increase inference steps from 64 to 256 (ablation) | CubiD-L · ImageNet 256x256 | gFID: 9.14 -> 5.33 | Improved |
| increase inference steps from 256 to 512 (ablation) | CubiD-L · ImageNet 256x256 | gFID: 5.33 -> 5.25 | No clear effect |
| scale model from 946M to 3.7B parameters (ablation) | CubiD · ImageNet 256x256 | gFID: 5.25 -> 4.68 | Improved |
| use SigLIP2 instead of DINOv2 as encoder (ablation) | CubiD-L · ImageNet 256x256 | gFID: 5.25 -> 5.87 | Hurt |
| use dimension-wise quantization instead of vector quantization on SigLIP2 features (ablation) | LLaVA · GQA | accuracy: 54.9 -> 63.1 | Improved |
| increase quantization levels L from 2 to 8 (ablation) | DINOv2-B dimension-wise quantization · ImageNet reconstruction | rFID: 1.38 -> 0.57 | Improved |
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.
- MaskGIT — MaskGIT: Masked Generative Image Transformer (CVPR 2022) · 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
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.88)
- Generative Multimodal Pretraining with Discrete Diffusion Timestep Tokens (CVPR 2025 · similarity 0.87)
- Discrete Modeling via Boundary Conditional Diffusion Processes (NeurIPS 2024 · similarity 0.87)
- ENAT: Rethinking Spatial-temporal Interactions in Token-based Image Synthesis (NeurIPS 2024 · similarity 0.87)
- Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction (CVPR 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.
