Research map / Image generation / Fast diffusion sampling
Transition Matching: Scalable and Flexible Generative Modeling
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.
Introduces Transition Matching, a discrete-time continuous-state generative framework generalizing diffusion/flow and continuous autoregressive models, with three variants (DTM, ARTM, FHTM) evaluated for text-to-image generation under a fixed DiT/LLM architecture and data.
Problem
Diffusion and flow matching design space is well explored, and continuous-token autoregressive models lack expressive, causal generation that matches flow-based quality; the paper seeks a unified, more flexible generative paradigm for media generation.
Approach
Generalize generative modeling to discrete-time Markov transitions learned by matching transition kernels under a supervising process; DTM learns the difference between endpoints (expressive stochastic kernel), while ARTM/FHTM use independent linear supervising processes with autoregressive, multi-step kernels trained with flow matching heads.
Key results
- DTM with 32 TM steps reaches PartiPrompts PickScore 21.18 and UnifiedReward 5.123, beating FM (21.04, 4.784) at 256 NFE, with DTM needing 32 backbone NFE.
- DTM sampling is about 7x faster than FM (1.6 s vs 10.8 s per image on H100) at comparable CLIPScore/PickScore.
- FHTM-3 (fully causal) reaches GenEval overall 0.52 and PartiPrompts CLIPScore 27.00, matching FM (GenEval 0.47) while being causal; ARTM-3 and FHTM-3 greatly improve over AR baseline (CLIPScore 24.85).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace flow matching with DTM | DiT image generation · PartiPrompts | PickScore: 21.04 -> 21.18 | Helped |
| Replace flow matching with DTM | DiT image generation · MS-COCO | PickScore: 21.11 -> 21.19 | Helped |
| Use DTM instead of FM | DiT image generation · PartiPrompts | time (sec): 10.8 -> 1.6 | Helped |
| Use DTM instead of FM | DiT image generation · PartiPrompts | CLIPScore: 26.0 -> 26.8 | Helped |
| Use ARTM-3 instead of AR baseline | DiT image generation · PartiPrompts | ImageReward: -0.428 -> 0.375 | Helped |
| Use FHTM-3 instead of FM | DiT image generation · GenEval | Overall: 0.47 -> 0.52 | Helped |
| Use MAR instead of DTM | DiT image generation · PartiPrompts | CLIPScore: 26.84 -> 26.98 | Hurt |
| Replace the dependent linear process with the independent linear process (ablation) | ARTM-3 and FHTM-3 · MS-COCO and PartiPrompts | CLIPScore and PickScore | Helped |
| Use an exponential rather than uniform transition scheduler (ablation) | ARTM-3, DiT · PartiPrompts | CLIPScore: 26.8494 -> 27.02 | Helped |
| Use an exponential rather than uniform transition scheduler (ablation) | FHTM-3, DiT · PartiPrompts | PickScore: 20.8797 -> 20.85 | No clear effect |
| Increase the DTM flow-head patch size from 2×2×4 to 2×8×4 (ablation) | DTM · PartiPrompts, low TM steps | CLIPScore and PickScore | Helped |
| Increase the DTM flow-head patch size from 2×2×4 to 2×8×4 (ablation) | DTM · PartiPrompts, high TM steps | CLIPScore | No clear effect |
Surprising findings
- Dependent linear supervising process causes AR kernels to learn a degenerate function and fail at inference, while the independent linear process is far superior.
- ARTM/FHTM gains come at the cost of more backbone forwards (TM steps times image tokens), offering no sampling speed-up.
Limitations
- ARTM/FHTM require backbone forwards equal to number of TM steps times number of image tokens, making sampling expensive.
- DTM generates each token independently within a step, limiting kernel expressiveness.
- Comparisons use only re-implemented baselines under fixed setup; no release statement.
Experimental setup
Datasets: Shutterstock image-caption pairs (350M, licensed)
Benchmarks: PartiPrompts MS-COCO GenEval
Metrics: CLIPScore PickScore ImageReward UnifiedReward Aesthetic DeQA Score GenEval overall
Baselines: FM MAR MAR-Fluid MAR-discrete AR AR-discrete FM-Restart
Builds on: flow matching MAR DiT Fluid
Closest papers by meaning
- Generator Matching: Generative modeling with arbitrary Markov processes (ICLR 2025 · similarity 0.90)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.88)
- Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective (ICLR 2025 · similarity 0.88)
- Inductive Moment Matching (ICML 2025 · similarity 0.87)
- Image Regeneration: Evaluating Text-to-Image Model via Generating Identical Image with Multimodal Large Language Models (AAAI 2025 · similarity 0.87)
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 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.