Pulse Explorar nav.journalClub Debates Tendencias Investigadores

Join discussions, follow papers, and never miss your next session.

Download on theApp Store

Inicio Explorar nav.journalClub Tendencias

⌘+K

SSD-Flow: Flow Matching Meets Mamba-2 for Scalable Image Generation | Synapse

February 21, 2026Open Access

Ver artículo completo

SSD-Flow: Flow Matching Meets Mamba-2 for Scalable Image Generation

Puntos clave

The aim is to create the first Mamba-2 model for image generation using flow matching techniques.
Developed SSD-Flow model combining Mamba-2 and flow matching for image generation.
Introduced innovative components such as Velocity-Conditioned SSD and Adaptive Timestep Routing.
Applied Logit-Normal State Scheduling to adapt the existing SSD framework.
Demonstrated effective integration of flow matching with image generation in SSD-Flow.
Showed improved scalability and image clarity compared to previous models.

Resumen

We propose SSD-Flow, the first Mamba-2 image generation model trained with flow matching. While LinGen (CVPR 2025) combined Mamba-2 with flow matching for video, no work has done so for pure image generation. SSD-Flow introduces Velocity-Conditioned SSD, Adaptive Timestep Routing, and Logit-Normal State Scheduling, adapting the SSD framework to flow matching ODE dynamics.

Preguntar a la IA

Me gusta

Guardar

Compartir

Ver artículo completo

Cite This Study

Hiroki Abe (Thu,) studied this question.

synapsesocial.com/papers/69994cd2873532290d0219b6 https://doi.org/https://doi.org/10.5281/zenodo.18697643

Preguntar a la IA

Me gusta

Guardar

Compartir

Ver artículo completo