Research map / Image generation / Fast diffusion sampling
Inverse Bridge Matching Distillation
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 Inverse Bridge Matching Distillation (IBMD), a universal distillation method that compresses conditional and unconditional diffusion bridge models into one-step or few-step generators using only corrupted inputs.
Problem
Diffusion bridge models (DBMs) for image-to-image translation require many sampling steps, and existing acceleration (consistency distillation, DBIM samplers) applies only to conditional DBMs and cannot yield one-step generators.
Approach
Formulate distillation as an inverse bridge matching problem: find a generator whose induced mixture of bridges yields the teacher's drift. This is turned into a tractable unconstrained objective (teacher loss minus inner bridge-matching loss) that needs only corrupted samples x_T and works for both conditional and unconditional bridges.
Key results
- 1-step IBMD-I2SB reaches FID 2.6 on 4x bicubic super-resolution vs 2.8 for the 1000-NFE I2SB teacher.
- On Edges->Handbags, 1-step IBMD-DDBM reaches FID 1.26 vs teacher's 0.89 at 100 NFE and CDBM CBT's 0.80 at 2 NFE; 2-step IBMD gets 0.67.
- For JPEG QF=10, 1-step IBMD-I2SB achieves FID 3.7 vs 3.6 for the 1000-NFE teacher, a ~100x speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill I²SB with IBMD into 1 NFE | I²SB · ImageNet 256×256, 4× super-resolution (bicubic) | FID: 2.8 -> 2.6 | Helped |
| distill I²SB with IBMD into 1 NFE | I²SB · ImageNet 256×256, JPEG restoration QF=5 | FID: 4.6 -> 5.2 | Hurt |
| distill I²SB with IBMD into 1 NFE | I²SB · ImageNet 256×256, 4× super-resolution (pool) | FID: 2.7 -> 2.5 | Helped |
| distill I²SB with IBMD into 1 NFE | I²SB · ImageNet 256×256, JPEG restoration QF=10 | FID: 3.6 -> 3.7 | Hurt |
| distill DDBM with IBMD into 2 NFE | DDBM · Edges → Handbags 64×64, image translation | FID: 0.89 -> 0.67 | Helped |
| distill DDBM with IBMD into 2 NFE | DDBM · DIODE-Outdoor 256×256, image translation | FID: 2.57 -> 3.12 | Hurt |
| distill DDBM with IBMD into 4 NFE | DDBM · ImageNet 256×256, center inpainting 128×128 | FID: 3.88 -> 4.03 | Hurt |
| use 2 NFE instead of 1 NFE (ablation) | IBMD-I²SB · ImageNet 256×256, center inpainting 128×128 | FID: 6.7 -> 5.3 | Helped |
| use 4 NFE instead of 1 NFE (ablation) | IBMD-I²SB · ImageNet 256×256, center inpainting 128×128 | FID: 6.7 -> 5.1 | Helped |
| use 2 NFE instead of 1 NFE (ablation) | IBMD-DDBM · ImageNet 256×256, center inpainting 128×128 | FID: 5.87 -> 4.23 | Helped |
| use 4 NFE instead of 1 NFE (ablation) | IBMD-DDBM · ImageNet 256×256, center inpainting 128×128 | FID: 5.87 -> 4.03 | Helped |
Surprising findings
- The one-step distilled generator outperforms its 1000-step teacher on 4x super-resolution, which the authors attribute to accumulated teacher sampling error.
Limitations
- Alternating between training an auxiliary bridge model and the student is computationally expensive.
- Student optimization requires backpropagation through teacher, auxiliary bridge and generator, about 3x the memory of training the teacher.
- One-step inpainting shows a noticeable quality gap versus the teacher.
Experimental setup
Datasets: ImageNet Edges->Handbags DIODE-Outdoor
Benchmarks: ImageNet 256x256 super-resolution (bicubic, pool) ImageNet JPEG restoration (QF=5, QF=10) ImageNet center inpainting 128x128 Edges->Handbags 64x64 DIODE-Outdoor 256x256
Metrics: FID Classifier Accuracy Inception Score
Baselines: DBIM CDBM (CBD) CDBM (CBT) I2SB DDRM Pix2Pix Palette DDNM PiGDM CDSB
Builds on: Bridge Matching I2SB DDBM DBIM DMD
Closest papers by meaning
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.91)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.90)
- Exploring the Design Space of Diffusion Bridge Models (NeurIPS 2025 · similarity 0.89)
- UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control (ICML 2025 · similarity 0.89)
- Consistency Diffusion Bridge Models (NeurIPS 2024 · similarity 0.89)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · 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.