SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI agentsEvoPresent

EvoPresent

EvoPresent is the official code repository for the paper "Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations" (ICLR2026), developed by the UCSB-AI team. It is a multi-agent framework for academic presentation (PPT/slide) generation that automatically performs content extraction, speech generation, storyline and script construction, image generation and knowledge retrieval augmentation, and slide design and rendering around paper content, and uses the built-in aesthetic checker PresAesth to evaluate and adjust the initial slides, achieving iterative self-improvement. The project also provides training infrastructure for the PresAesth aesthetic model, presentation video generation capabilities (including lip-sync animation and TTS configuration), and the EvoPresent Benchmark evaluation dataset, making it suitable for researchers and developers to explore automated, high-quality presentation generation solutions in scenarios such as academic reporting and paper explanation.

Open website

About this tool

Project Positioning

EvoPresent is the official code repository for the paper "Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations", labeled ICLR2026 and released by the UCSB-AI team. The project focuses on the automatic generation and optimization of academic presentations, adopting the approach of Self-Improvement Aesthetic Agents.

Main Capabilities

According to the project description, the EvoPresent agent workflow includes the following stages:

  • Content extraction and speech generation;
  • Construction of storylines and scripts;
  • Content enhancement using image generation and knowledge retrieval;
  • Design and rendering of slides;
  • Evaluation of initial slides by the aesthetic checker with adjustment suggestions;
  • Iterative self-improvement of the agent framework guided by the PresAesth model.

PresAesth is fine-tuned based on Qwen2.5-VL-7B-Instruct for presentation aesthetic evaluation tasks. The training tasks include scoring, defect adjustment, and comparison, and the training data is a human preference aesthetic dataset.

Usage

The project provides installation and inference guidance:

  • For API setup, gpt-4o-2024-08-06 and gpt-4.1-2025-04-14 are recommended for the text and image extraction stages; for the slide generation stage, deepseek-chat/deepseek-reasoner, gemini-2.5-pro, and claude-sonnet-4-20250514 are recommended.
  • In evopresent/ppt/gen_pptx_code.py, a temperature of 0.6–0.8 is recommended and can be adjusted for different models.
  • Quick inference: create a folder named {paper_name} under {dataset_dir} and place the paper in it as paper.pdf.
  • checker-scope controls the layout check scope: all (check all slides), images (check only slides containing images/tables), text (check only plain textslides). style specifies the presentation theme and visual style (such as color scheme, fonts, whitespace, animations, etc.); for example, tech_dark is a dark tech style. Evopresent/evopresent/styles provides a variety of presentation style templates.
  • In the presentation video generation section, the parameters related to input and output directories include: --html-dir (directory of presentation HTML files), --script-json (JSON file of scripts for each page), --ref-face (path to the reference face image used to generate facial animation), --ckpt-path (model weight file, such as float.pth), --output-dir (audio and video output directory), --final-video (save path of the final composited video).
  • TTS configuration parameters include: --tts-backend (choose between megatts3 and OpenAI), --openai-api-key (required when using OpenAI TTS), --openai-voice and --openai-speed (voice and speech rate parameters for OpenAI TTS), --voice-wav (reference WAV file used when choosing the megatts3 backend).
  • Video parameters include: --frame-width and --frame-height (size of each slide page in the video), --margin-x and --margin-y (margins for the face position in the video frame).

Presentation Generation Options and Training

Regarding TTS, the project description states that if speed is desired, OpenAI TTS is recommended, and for voice selection please refer to the official OpenAI documentation; if personalized voice is needed, you can use MegaTTS3's voice cloning tool to imitate your own voice, or view several preset voices in the Google Drive folder.

The PresAesth training section includes infrastructure for training an aesthetics evaluation model based on Qwen2.5-VL-7B-Instruct. The commands include: first install uv (pip install uv), enter the train directory and execute bash ./setup.sh; configure the API key by editing the .env file (cp env-template .env); during training execute source train_env/bin/activate and bash ./run.sh; for evaluation and inference execute python eval.py and python inference.py.

Evaluation Dataset

The project has released the EvoPresent Benchmark evaluation dataset (hosted on Hugging Face). Due to potential copyright restrictions, the images cannot be redistributed directly, but they can be accessed through the links provided in the metadata. The project also mentions that it references the code and data of open-r1-multimodal, Paper2Poster, Float, and MegaTTS3, and provides citation information.

Applicable Scenarios and Limitations

The applicable scenarios are mainly automatic generation of academic presentations, production of paper explanation videos, and training and research of presentation aesthetics evaluation models. In terms of limitations, the project relies on external large model APIs (such as OpenAI, DeepSeek, Gemini, Claude) and TTS services, requiring corresponding API keys; the video generation part involves additional environment configuration for components such as FLOAT and MegaTTS3; the images in the evaluation dataset are subject to copyright restrictions and cannot be redistributed directly. Other features, pricing, or platform information not described in the materials are not provided for now.

Related tools

Browse category