This paper introduces ComfyMind, a novel collaborative AI system for general-purpose generative tasks leveraging ComfyUI.

dataset for multi-modal learning

An LLM-based framework that automates executable environment building to enable scaling up executable code data.

We introduce CaPT, leveraging both instance-aware and task-aware information for effective and efficient prompt-based learning.

A simple yet powerful Diffusion ConvNet

We propose a training-free visual token pruning method CDPruner for MLLM inference acceleration by maximizing the conditional diversity of retained tokens.

We propose a novel model ``MoE-ization'' strategy using SVD, which leads to a conflict- and oblivion-resistant multi-task adaptation method.

This paper introduces a unified visual tokenizer to facilitate unification of visual generation and understanding within a single autoregressive framework.

Efficient Reasoning VLM

We propose a more expressive Transformer, which exceeding its original $TC^0$ expressiveness.

PartCrafter is a structured 3D generative model that jointly generates multiple parts and objects from a single RGB image in one shot.

We proposed a new data selection method for pretraining multilingual Large Language Models

We propose the Multi-Reward Optimization (MRO) approach, which enhances token correlation during the denoising process in diffusion language models, improving reasoning performance and sampling efficiency.

InfinityStar is the first discrete autoregressive video generator capable of producing industrial-level 720p videos.

We propose SuperCLIP, a simple and efficient extension to CLIP that adds classfication-based supervision to improve fine-grained image-text alignment without requiring extra annotations or significant computation.

We propose a data selection method that leverages sparse, monosemantic neuronal activations learned via a sparse autoencoder to improve task-specific instruction tuning for large language models.

REArtGS

This paper proposes the ACT data pipeline, which reduces human annotation costs by using MLLMs as annotators and error detectors, and provides a theoretical analysis to ensure effective downstream training.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks.