Research map / Image generation / Fast diffusion sampling
BADiff: Bandwidth Adaptive Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BADiff, a diffusion model conditioned on a target entropy (bitrate) budget with an entropy hinge loss and adaptive stop policy, so images are generated directly at a bandwidth-matched bit-rate instead of generate-then-compress.
Problem
Cloud-to-device image delivery under limited bandwidth: diffusion models use a fixed number of steps and rely on post-generation lossy compression, which wastes computation and degrades fine textures.
Approach
Condition every denoising step on a target entropy budget via a learned embedding (FiLM), penalize the differentiable neural entropy estimate of the predicted clean image above that budget, and learn a stop-probability policy to terminate sampling early.
Key results
- On CIFAR-10 with DDPM-1k at low bitrate, BADiff reports FID 11.4 versus 13.6 for DDPM+LIC and 15.2 for DDPM+BPG.
- On CIFAR-10 at low bitrate, BADiff inference is 65 ms versus 115 ms for Cascade+LIC on DDPM-1k (reported 1.7x speed-up).
- Text-to-image on Stable Diffusion at low bitrate: BADiff FID 26.1 versus 30.7 for Cascade (SD+LIC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use entropy-conditioned BADiff instead of DDPM + LIC | DDPM-1k · CIFAR-10, low bitrate | FID: 13.6 -> 11.4 | Helped |
| Use entropy-conditioned BADiff instead of DDPM + LIC | DDPM-1k · CelebA-HQ, low bitrate | FID: 25.3 -> 21.7 | Helped |
| Use entropy-conditioned BADiff instead of DDPM + LIC | DDPM-1k · LSUN, low bitrate | FID: 22.8 -> 19.6 | Helped |
| Remove entropy conditioning (ablation) | BADiff · CIFAR-10, low bitrate | FID: 11.4 -> 13.1 (+1.7) | Hurt |
| Remove entropy conditioning (ablation) | BADiff · CIFAR-10, low bitrate | Δ bpp: 0.021 -> 0.038 (+0.017 bpp) | Hurt |
| Remove entropy hinge loss (ablation) | BADiff · CIFAR-10, low bitrate | FID: 11.4 -> 16.2 | Hurt |
| Remove entropy hinge loss (ablation) | BADiff · CIFAR-10, low bitrate | Δ bpp: 0.021 -> 0.055 | Hurt |
| Remove calibration loss (ablation) | BADiff · CIFAR-10, low bitrate | FID: 11.4 -> 18.6 (+7.2) | Hurt |
| Remove calibration loss (ablation) | BADiff · CIFAR-10, low bitrate | Δ bpp: 0.021 -> 0.043 (+0.022 bpp) | Hurt |
| Remove entropy conditioning (ablation) | BADiff · CIFAR-10, low bitrate | Time (ms): 65 -> 64 | No clear effect |
| Remove entropy hinge loss (ablation) | BADiff · CIFAR-10, low bitrate | Time (ms): 65 -> 65 | No clear effect |
| Remove calibration loss (ablation) | BADiff · CIFAR-10, low bitrate | Time (ms): 65 -> 65 | No clear effect |
Surprising findings
- Naive early-stop with LIC yields much worse FID than full-step cascades, despite matching bitrate.
- Claims of LPIPS and FID being held constant while stopping about 50% earlier are asserted in text without matching tables.
Limitations
- Targets spatially uniform entropy budgets only
- Evaluated only up to 256^2 resolution in the main paper (high-resolution experiments are limited)
- Stated entropy predictor is an approximation of codec bitrate; bpp adherence depends on it
Experimental setup
Datasets: CIFAR-10 CelebA-HQ LSUN-Church LSUN-Bedroom COCO-val2017
Benchmarks: CIFAR-10 CelebA-HQ LSUN
Metrics: FID LPIPS bpp inference time (ms)
Baselines: DDPM + BPG DDPM + LIC Early-Stop + LIC PNDM + LIC DPM-Solver + LIC LDM + BPG LDM + LIC
Builds on: DDPM latent diffusion models DDIM FiLM Cheng2020 learned image compression
Closest papers by meaning
- Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy (NeurIPS 2024 · similarity 0.88)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.88)
- AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.87)
- Schedule On the Fly: Diffusion Time Prediction for Faster and Better Image Generation (CVPR 2025 · similarity 0.86)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.86)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.86)
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.