Distribution Matching Distillation
PulseAugur coverage of Distribution Matching Distillation — every cluster mentioning Distribution Matching Distillation across labs, papers, and developer communities, ranked by signal.
7 day(s) with sentiment data
-
New eDMD method improves diffusion model distillation for faster generation
Researchers have developed Endpoint Distribution Matching Distillation (eDMD), a novel method for distilling diffusion models into fewer steps. Unlike previous techniques that use intermediate student outputs, eDMD trai…
-
APEX improves video generation with condition shifting, reducing costs
Researchers have developed APEX, a new method for few-step streaming video generation that improves upon existing techniques by using condition shifting. This approach eliminates the need for a separate fake-score netwo…
-
New distillation methods boost discrete diffusion model efficiency and quality
Researchers have developed new methods for distilling discrete diffusion models to improve efficiency and quality. One approach, Bidirectional Centered Score Matching (BCSM), offers an efficient objective that reduces c…
-
New distillation method improves AI model training with joint scaling and step-aware critics
Researchers have developed a new method called Distribution Matching Distillation (DMD) to improve the training of few-step student models. By analyzing the critic's capacity and information, they found that scaling the…
-
New IDM technique enhances text-to-image generation accuracy
Researchers have introduced Intrinsic Distribution Matching (IDM), a novel technique for text-to-image generation that aims to improve prompt adherence with limited refinement steps. IDM decomposes the reverse-KL field …
-
New IDM technique boosts text-to-image generation with moment constraints
Researchers have developed Intrinsic Distribution Matching (IDM), a novel technique to enhance few-step text-to-image generation. IDM incorporates explicit moment constraints into distribution matching distillation, ena…
-
New technique enhances few-step video generation with representation forcing
Researchers have introduced Self Flow Forcing (SF$^2$), a novel method to improve few-step video generation by incorporating representation learning into distribution matching distillation. This approach addresses limit…
-
AdvFidelity integrates preference alignment and few-step image generation
Researchers have introduced AdvFidelity, a novel approach that unifies preference alignment and efficient few-step image generation within a single reinforcement learning (RL) stage. This method integrates artifact-sens…
-
New data-space iteration method improves generative model sampling
Researchers have developed a new few-step generation framework called data-space iteration, which removes the need for flow discretization in generative models. This method allows a shared generator to directly refine i…
-
New MBDMD method bridges Drifting Models and Distribution Matching Distillation
This paper introduces Multi-Bandwidth Distribution Matching Distillation (MBDMD), an advancement on Distribution Matching Distillation (DMD). Researchers have established a connection between Diffusion & Flow Style Gene…
-
New data-space iteration method speeds up generative model sampling
Researchers have developed a new few-step generation framework called data-space iteration, which eliminates the need for flow discretization in generative models. This method allows a shared generator to directly refin…
-
New DMAD method accelerates visual generation via adversarial distillation
Researchers have introduced DMAD, a novel method for faster visual generation that reframes distribution matching as a classification problem. This approach trains a student model without the need for auxiliary diffusio…
-
DMAD accelerates visual generation with adversarial distillation
Researchers have developed DMAD, a novel method for adversarial distillation that significantly speeds up visual generation processes. Unlike previous techniques that require auxiliary diffusion models, DMAD uses discri…
-
DyMD framework enhances few-step video generation by preserving interaction dynamics
Researchers have developed DyMD, a novel framework for distilling large video diffusion models into smaller, faster ones. While existing methods like Distribution Matching Distillation (DMD) can generate videos in fewer…
-
AI research tackles long-video memory with novel weight-internalization techniques · 7 sources tracked
Researchers are developing novel methods to enhance the memory capabilities of AI models, particularly for processing long video streams. Several papers introduce techniques that internalize historical data into model w…
-
New methods enhance autoregressive video generation by tackling mode collapse and diversity loss
Researchers have developed two new methods to improve autoregressive video generation models, addressing issues of mode collapse and diversity loss. The first method, Mask Forcing, injects masked cleaner signals during …
-
DynaForcing framework tackles dynamic collapse in avatar generation
Researchers have introduced DynaForcing, a new training framework designed to improve audio-driven avatar generation by addressing the issue of dynamic collapse. This phenomenon causes student models to produce static o…
-
SQuad framework slashes Video Transformer compute costs with sub-quadratic attention
Researchers have developed SQuad, a Sub-Quadratic Attention Distillation framework designed to improve the efficiency of Video Diffusion Transformers (DiTs). This new method reduces the computational cost of the self-at…
-
DistillAlign improves video distillation via distributional alignment
Researchers have introduced DistillAlign, a novel approach to autoregressive video distillation that addresses limitations in existing multi-stage pipelines. The method emphasizes distributional alignment between studen…
-
AudioX-Turbo framework enables efficient multimodal audio generation
Researchers have introduced AudioX-Turbo, a novel framework designed for efficient generation of audio from various multimodal inputs like text, video, and audio signals. The system employs a teacher-student distillatio…