Tag
This paper introduces Wiener-Informed Decoding (WiD), which leverages analytic Wiener-filter-style velocity estimators derived from second-order data statistics to ease training of pixel-space flow models, initializing new pathways within pixel decoders that are jointly trained for improved class-conditional and text-to-image generation across resolutions.
The paper presents Lafa, an 8B-parameter blockwise continuous flow language model trained with under 30B tokens that matches or surpasses autoregressive quality while achieving 2.54x inference throughput via self-speculative decoding and blockwise parallel decoding.
论文提出 LLM-to-Flow Distillation（LFD），利用冻结的 LLM 作为自适应教师构建课程并通过 DPO 将偏好蒸馏到科学生成器中，无需显式代理奖励即可用自然语言目标控制治疗性肽和小分子结构的生成设计。
A framework combining deep generative models (flow and score-based) with symbolic regression to automatically infer analytical forms of unnormalized probability densities from samples, demonstrated on toy distributions and lattice physics systems like the XY model and φ⁴ theory renormalization.
SpaceFlow 提出了一种免训练的局部可控 3D 生成流程，用户可通过几何基元为不同区域指定控制强度，并分别路由文本或图像提示以合成外观，从而在几何保真与生成自由度之间取得平衡。
The tweet discusses the growing use of continuous flow models for discrete tasks like language generation, proposing a method to represent vocabulary as one-hot vectors and learn flows in continuous space.
This paper introduces Newton Matching, a unified framework for fine-tuning and sampling in generative models, which addresses limitations of existing methods by treating learning as an iterative optimization process and leveraging conditional-matching structure.
PathGuide reformulates classifier-free guidance selection as an on-policy transport problem in flow-based generative models, using the weak form of the continuity equation to dynamically optimize guidance scales for improved sample fidelity.
The article introduces ActFlow, a continued pre-training scheme that expands the valid design space for flow and diffusion models, enabling out-of-distribution generative modeling and evolvable search spaces in scientific discovery.
This paper proposes MeanFlow-Transfer (MF-T) and Continuous Adversarial MeanFlow (CAMF) to unify the adaptation and acceleration of pretrained diffusion and flow models, enabling high-quality few-step generation on new domains with limited data.
GS-Voxel introduces a fitting-free framework to convert 3D Gaussian Splatting reconstructions into structured latents, enabling scalable generation of large-scale aerial 3D scenes via flow models and tiled inference.
Introduces CrystalGRPO, a reinforcement-learning post-training framework for flow-based crystal structure prediction that aligns target recovery and preserves candidate coverage, improving Top-1 and Top-20 performance across MP-20 and MPTS-52 benchmarks.
This paper introduces LC-GRPO, a flow-based GRPO framework with Langevin correction that bridges the train-inference gap by aligning stochastic training rollouts with deterministic ODE sampling, improving reward optimization on models like SD3.5, FLUX.1-Dev, and HunyuanVideo.
FLUX 3 proposes multimodal flow models as a foundational approach for real-world visual intelligence, building on prior FLUX work.
MeanFlowNFT introduces a forward-process reinforcement learning method for average-velocity generators, enabling efficient alignment with human preferences while preserving fast few-step sampling. Experiments show it outperforms prior RL-tuned few-step generators on most metrics and even surpasses multi-step RL-tuned diffusion models.
Flow Reasoning Models (FRMs) introduce a training and test-time-scaling framework for discrete flow models on structured reasoning tasks. By using self-verification and self-conditioning, FRMs achieve nearly 100% solve rates on Sudoku and Zebra puzzles with far fewer passes than previous baselines.
This paper introduces Masked Language Flow Models (MLFMs), which incorporate masking into flow-based language models to enable continuous flow for conditional generation and allow pretrained Masked Diffusion Models to be converted. The authors propose a novel sampler that alternates continuous denoising with discrete unmasking, demonstrating for the first time that flow-based language models can scale to downstream reasoning and instruction-following tasks.
FlowBender is a closed-loop framework that improves constraint satisfaction in diffusion and flow models by training networks to correct alignment errors using inference-time feedback, outperforming traditional supervised and guidance-based approaches.
Proposes treating flow steps as RL actions combined with a 'flow reversal' technique for flow offline reinforcement learning.
QGF is an RL algorithm that improves policies at test time by using a value gradient to guide a pre-trained flow policy, avoiding training-time instability while maintaining competitive performance.