LL3M: Large Language 3D Modelers

We present LL3M, a multi-agent system that leverages pretrained large language models (LLMs) to generate 3D assets by writing interpretable Python code in Blender. We break away from the typical generative approach that learns from a collection of 3D data. Instead, we reformulate shape generation as a code-writing task, enabling greater modularity, editability, and integration with artist workflows. Given a text prompt, LL3M coordinates a team of specialized LLM agents to plan, retrieve, write, debug, and refine Blender scripts that generate and edit geometry and appearance. The generated code works as a high-level, interpretable, human-readable, well-documented representation of scenes and objects, making full use of sophisticated Blender constructs (e.g. B-meshes, geometry modifiers, shader nodes) for diverse, unconstrained shapes, materials, and scenes. This code presents many avenues for further agent and human editing and experimentation via code tweaks or procedural parameters. This medium naturally enables a co-creative loop in our system: agents can automatically self-critique using code and visuals, while iterative user instructions provide an intuitive way to refine assets. A shared code context across agents enables awareness of previous attempts, and a retrieval-augmented generation knowledge base built from Blender API documentation, BlenderRAG, equips agents with examples, types, and functions empowering advanced modeling operations and code correctness. We demonstrate the effectiveness of LL3M across diverse shape categories, style and material edits, and user-driven refinements. Our experiments showcase the power of code as a generative and interpretable medium for 3D asset creation. Our project page is at https://threedle.github.io/ll3m.

MeshCoder: LLM-Powered Structured Mesh Code Generation from Point Clouds

Reconstructing 3D objects into editable programs is pivotal for applications like reverse engineering and shape editing. However, existing methods often rely on limited domain-specific languages (DSLs) and small-scale datasets, restricting their ability to model complex geometries and structures. To address these challenges, we introduce MeshCoder, a novel framework that reconstructs complex 3D objects from point clouds into editable Blender Python scripts. We develop a comprehensive set of expressive Blender Python APIs capable of synthesizing intricate geometries. Leveraging these APIs, we construct a large-scale paired object-code dataset, where the code for each object is decomposed into distinct semantic parts. Subsequently, we train a multimodal large language model (LLM) that translates 3D point cloud into executable Blender Python scripts. Our approach not only achieves superior performance in shape-to-code reconstruction tasks but also facilitates intuitive geometric and topological editing through convenient code modifications. Furthermore, our code-based representation enhances the reasoning capabilities of LLMs in 3D shape understanding tasks. Together, these contributions establish MeshCoder as a powerful and flexible solution for programmatic 3D shape reconstruction and understanding.

MapStory: LLM-Powered Text-Driven Map Animation Prototyping with Human-in-the-Loop Editing

We introduce MapStory, an LLM-powered animation authoring tool that generates editable map animation sequences directly from natural language text. Given a user-written script, MapStory leverages an agentic architecture to automatically produce a scene breakdown, which decomposes the script into key animation building blocks such as camera movements, visual highlights, and animated elements. Our system includes a researcher component that accurately queries geospatial information by leveraging an LLM with web search, enabling the automatic extraction of relevant regions, paths, and coordinates while allowing users to edit and query for changes or additional information to refine the results. Additionally, users can fine-tune parameters of these blocks through an interactive timeline editor. We detail the system's design and architecture, informed by formative interviews with professional animators and an analysis of 200 existing map animation videos. Our evaluation, which includes expert interviews (N=5) and a usability study (N=12), demonstrates that MapStory enables users to create map animations with ease, facilitates faster iteration, encourages creative exploration, and lowers barriers to creating map-centric stories.

CAD-Coder:Text-Guided CAD Files Code Generation

Computer-aided design (CAD) is a way to digitally create 2D drawings and 3D models of real-world products. Traditional CAD typically relies on hand-drawing by experts or modifications of existing library files, which doesn't allow for rapid personalization. With the emergence of generative artificial intelligence, convenient and efficient personalized CAD generation has become possible. However, existing generative methods typically produce outputs that lack interactive editability and geometric annotations, limiting their practical applications in manufacturing. To enable interactive generative CAD, we propose CAD-Coder, a framework that transforms natural language instructions into CAD script codes, which can be executed in Python environments to generate human-editable CAD files (.Dxf). To facilitate the generation of editable CAD sketches with annotation information, we construct a comprehensive dataset comprising 29,130 Dxf files with their corresponding script codes, where each sketch preserves both editability and geometric annotations. We evaluate CAD-Coder on various 2D/3D CAD generation tasks against existing methods, demonstrating superior interactive capabilities while uniquely providing editable sketches with geometric annotations.

REGen: Multimodal Retrieval-Embedded Generation for Long-to-Short Video Editing

Short videos are an effective tool for promoting contents and improving knowledge accessibility. While existing extractive video summarization methods struggle to produce a coherent narrative, existing abstractive methods cannot `quote' from the input videos, i.e., inserting short video clips in their outputs. In this work, we explore novel video editing models for generating shorts that feature a coherent narrative with embedded video insertions extracted from a long input video. We propose a novel retrieval-embedded generation framework that allows a large language model to quote multimodal resources while maintaining a coherent narrative. Our proposed REGen system first generates the output story script with quote placeholders using a finetuned large language model, and then uses a novel retrieval model to replace the quote placeholders by selecting a video clip that best supports the narrative from a pool of candidate quotable video clips. We examine the proposed method on the task of documentary teaser generation, where short interview insertions are commonly used to support the narrative of a documentary. Our objective evaluations show that the proposed method can effectively insert short video clips while maintaining a coherent narrative. In a subjective survey, we show that our proposed method outperforms existing abstractive and extractive approaches in terms of coherence, alignment, and realism in teaser generation.

Talk to Your Slides: Efficient Slide Editing Agent with Large Language Models

Existing research on large language models (LLMs) for PowerPoint predominantly focuses on slide generation, overlooking the common yet tedious task of editing existing slides. We introduce Talk-to-Your-Slides, an LLM-powered agent that directly edits slides within active PowerPoint sessions through COM communication. Our system employs a two-level approach: (1) high-level processing where an LLM agent interprets instructions and formulates editing plans, and (2) low-level execution where Python scripts directly manipulate PowerPoint objects. Unlike previous methods relying on predefined operations, our approach enables more flexible and contextually-aware editing. To facilitate evaluation, we present TSBench, a human-annotated dataset of 379 diverse editing instructions with corresponding slide variations. Experimental results demonstrate that Talk-to-Your-Slides significantly outperforms baseline methods in execution success rate, instruction fidelity, and editing efficiency. Our code and benchmark are available at https://anonymous.4open.science/r/talk-to-your-slides/

VC-LLM: Automated Advertisement Video Creation from Raw Footage using Multi-modal LLMs

As short videos have risen in popularity, the role of video content in advertising has become increasingly significant. Typically, advertisers record a large amount of raw footage about the product and then create numerous different short-form advertisement videos based on this raw footage. Creating such videos mainly involves editing raw footage and writing advertisement scripts, which requires a certain level of creative ability. It is usually challenging to create many different video contents for the same product, and manual efficiency is often low. In this paper, we present VC-LLM, a framework powered by Large Language Models for the automatic creation of high-quality short-form advertisement videos. Our approach leverages high-resolution spatial input and low-resolution temporal input to represent video clips more effectively, capturing both fine-grained visual details and broader temporal dynamics. In addition, during training, we incorporate supplementary information generated by rewriting the ground truth text, ensuring that all key output information can be directly traced back to the input, thereby reducing model hallucinations. We also designed a benchmark to evaluate the quality of the created videos. Experiments show that VC-LLM based on GPT-4o can produce videos comparable to those created by humans. Furthermore, we collected numerous high-quality short advertisement videos to create a pre-training dataset and manually cleaned a portion of the data to construct a high-quality fine-tuning dataset. Experiments indicate that, on the benchmark, the VC-LLM based on fine-tuned LLM can produce videos with superior narrative logic compared to those created by the VC-LLM based on GPT-4o.

CodeMonkeys: Scaling Test-Time Compute for Software Engineering

Scaling test-time compute is a promising axis for improving LLM capabilities. However, test-time compute can be scaled in a variety of ways, and effectively combining different approaches remains an active area of research. Here, we explore this problem in the context of solving real-world GitHub issues from the SWE-bench dataset. Our system, named CodeMonkeys, allows models to iteratively edit a codebase by jointly generating and running a testing script alongside their draft edit. We sample many of these multi-turn trajectories for every issue to generate a collection of candidate edits. This approach lets us scale "serial" test-time compute by increasing the number of iterations per trajectory and "parallel" test-time compute by increasing the number of trajectories per problem. With parallel scaling, we can amortize up-front costs across multiple downstream samples, allowing us to identify relevant codebase context using the simple method of letting an LLM read every file. In order to select between candidate edits, we combine voting using model-generated tests with a final multi-turn trajectory dedicated to selection. Overall, CodeMonkeys resolves 57.4% of issues from SWE-bench Verified using a budget of approximately 2300 USD. Our selection method can also be used to combine candidates from different sources. Selecting over an ensemble of edits from existing top SWE-bench Verified submissions obtains a score of 66.2% and outperforms the best member of the ensemble on its own. We fully release our code and data at https://scalingintelligence.stanford.edu/pubs/codemonkeys.

Lotus: Creating Short Videos From Long Videos With Abstractive and Extractive Summarization

Short-form videos are popular on platforms like TikTok and Instagram as they quickly capture viewers' attention. Many creators repurpose their long-form videos to produce short-form videos, but creators report that planning, extracting, and arranging clips from long-form videos is challenging. Currently, creators make extractive short-form videos composed of existing long-form video clips or abstractive short-form videos by adding newly recorded narration to visuals. While extractive videos maintain the original connection between audio and visuals, abstractive videos offer flexibility in selecting content to be included in a shorter time. We present Lotus, a system that combines both approaches to balance preserving the original content with flexibility over the content. Lotus first creates an abstractive short-form video by generating both a short-form script and its corresponding speech, then matching long-form video clips to the generated narration. Creators can then add extractive clips with an automated method or Lotus's editing interface. Lotus's interface can be used to further refine the short-form video. We compare short-form videos generated by Lotus with those using an extractive baseline method. In our user study, we compare creating short-form videos using Lotus to participants' existing practice.

IMAGINE-E: Image Generation Intelligence Evaluation of State-of-the-art Text-to-Image Models

With the rapid development of diffusion models, text-to-image(T2I) models have made significant progress, showcasing impressive abilities in prompt following and image generation. Recently launched models such as FLUX.1 and Ideogram2.0, along with others like Dall-E3 and Stable Diffusion 3, have demonstrated exceptional performance across various complex tasks, raising questions about whether T2I models are moving towards general-purpose applicability. Beyond traditional image generation, these models exhibit capabilities across a range of fields, including controllable generation, image editing, video, audio, 3D, and motion generation, as well as computer vision tasks like semantic segmentation and depth estimation. However, current evaluation frameworks are insufficient to comprehensively assess these models' performance across expanding domains. To thoroughly evaluate these models, we developed the IMAGINE-E and tested six prominent models: FLUX.1, Ideogram2.0, Midjourney, Dall-E3, Stable Diffusion 3, and Jimeng. Our evaluation is divided into five key domains: structured output generation, realism, and physical consistency, specific domain generation, challenging scenario generation, and multi-style creation tasks. This comprehensive assessment highlights each model's strengths and limitations, particularly the outstanding performance of FLUX.1 and Ideogram2.0 in structured and specific domain tasks, underscoring the expanding applications and potential of T2I models as foundational AI tools. This study provides valuable insights into the current state and future trajectory of T2I models as they evolve towards general-purpose usability. Evaluation scripts will be released at https://github.com/jylei16/Imagine-e.