Research map / Probabilistic and scientific modeling / Variational and generative sampling
IDF++: Analyzing and Improving Integer Discrete Flows for Lossless Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply the IDF++ architecture modifications | IDF · CIFAR-10 | bits per dimension (compression): 3.32 -> 3.26 | Helped |
| Apply the IDF++ architecture modifications | IDF · ImageNet-32 | bits per dimension (compression): 4.18 -> 4.12 | Helped |
| Apply the IDF++ architecture modifications | IDF · ImageNet-64 | bits per dimension (compression): 3.90 -> 3.81 | Helped |
| Use 4 instead of 8 flows per level | IDF++ · CIFAR-10 | bits per dimension (compression): 3.26 -> 3.31 | Mixed |
| Use 4 instead of 8 flows per level | IDF++ · ImageNet-32 | bits per dimension (compression): 4.12 -> 4.16 | Mixed |
| Use 4 instead of 8 flows per level | IDF++ · ImageNet-64 | bits per dimension (compression): 3.81 -> 3.85 | Mixed |
| Train on 100% rather than 80% of the training data (ablation) | IDF · CIFAR-10 | theoretically achievable bits per dimension: 3.322 -> 3.298 | Helped |
| Add EMA weights (ablation) | IDF · CIFAR-10 | theoretically achievable bits per dimension: 3.298 -> 3.291 | Helped |
| Add rezero and double the learning rate (ablation) | IDF + EMA · CIFAR-10 | theoretically achievable bits per dimension: 3.291 -> 3.262 | Helped |
| Invert channel permutations (ablation) | IDF + EMA + rezero + lr × 2 · CIFAR-10 | theoretically achievable bits per dimension: 3.262 -> 3.255 | Helped |
| Add group normalization and Swish activations (ablation) | IDF + EMA + rezero + lr × 2 + invert perm · CIFAR-10 | theoretically achievable bits per dimension: 3.255 -> 3.241 | Helped |
| Replace the identity straight-through estimator with less-biased alternatives, including soft rounding | IDF · CIFAR-10 | bits per dimension | No clear effect |
Closest papers by meaning
- Integer Discrete Flows and Lossless Compression (NeurIPS 2019 · similarity 0.93)
- Fast Lossless Neural Compression with Integer-Only Discrete Flows (ICML 2022 · similarity 0.90)
- iFlow: Numerically Invertible Flows for Efficient Lossless Compression via a Uniform Coder (NeurIPS 2021 · similarity 0.88)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · similarity 0.85)
- LotteryCodec: Searching the Implicit Representation in a Random Network for Low-Complexity Image Compression (ICML 2025 · similarity 0.84)
- Practical Shuffle Coding (NeurIPS 2024 · similarity 0.84)
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.