Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2511.11418 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866917080304975872 |
|---|---|
| author | Varam, Dara Abuhani, Diaa A. Zualkernan, Imran AlDamani, Raghad Khalil, Lujain |
| author_facet | Varam, Dara Abuhani, Diaa A. Zualkernan, Imran AlDamani, Raghad Khalil, Lujain |
| contents | Flow Matching (FM) generative models offer efficient simulation-free training and deterministic sampling, but their practical deployment is challenged by high-precision parameter requirements. We adapt optimal transport (OT)-based post-training quantization to FM models, minimizing the 2-Wasserstein distance between quantized and original weights, and systematically compare its effectiveness against uniform, piecewise, and logarithmic quantization schemes. Our theoretical analysis provides upper bounds on generative degradation under quantization, and empirical results across five benchmark datasets of varying complexity show that OT-based quantization preserves both visual generation quality and latent space stability down to 2-3 bits per parameter, where alternative methods fail. This establishes OT-based quantization as a principled, effective approach to compress FM generative models for edge and embedded AI applications. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_11418 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Low-Bit, High-Fidelity: Optimal Transport Quantization for Flow Matching Varam, Dara Abuhani, Diaa A. Zualkernan, Imran AlDamani, Raghad Khalil, Lujain Machine Learning Computer Vision and Pattern Recognition Flow Matching (FM) generative models offer efficient simulation-free training and deterministic sampling, but their practical deployment is challenged by high-precision parameter requirements. We adapt optimal transport (OT)-based post-training quantization to FM models, minimizing the 2-Wasserstein distance between quantized and original weights, and systematically compare its effectiveness against uniform, piecewise, and logarithmic quantization schemes. Our theoretical analysis provides upper bounds on generative degradation under quantization, and empirical results across five benchmark datasets of varying complexity show that OT-based quantization preserves both visual generation quality and latent space stability down to 2-3 bits per parameter, where alternative methods fail. This establishes OT-based quantization as a principled, effective approach to compress FM generative models for edge and embedded AI applications. |
| title | Low-Bit, High-Fidelity: Optimal Transport Quantization for Flow Matching |
| topic | Machine Learning Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2511.11418 |