VideTShrts

AI video research, in plain English

The ideas inside the tools: how clippers find the best moments, how auto captions work, and how video generators are built. One paragraph per paper, with a link to the original.

Open video models

Wan: Open and Advanced Large-Scale Video Generative Models

arXiv, March 26, 2025

Alibaba's report on the open Wan video models, including a small version meant to run on consumer GPUs.

Why it matters: Wan's open releases became the default local video model; its closed successor Wan 3.0 now competes with the top commercial models.

LTX-Video: Realtime Video Latent Diffusion

arXiv, December 30, 2024

Lightricks' open model built for speed: it aims to generate video faster than real time on high-end GPUs.

Why it matters: Speed is what makes generation usable for quick iterations and drafts, and it is the base of the later open LTX-2 releases.

Video generation

Movie Gen: A Cast of Media Foundation Models

arXiv, October 17, 2024

Meta's technical report on a family of models that generate video with synced audio, personalize video from a photo and edit video from text.

Why it matters: One of the most detailed public write-ups of how large video generators are built and evaluated, useful for reading vendor claims critically.

Finding the best moments

UniVTG: Towards Unified Video-Language Temporal Grounding

arXiv, July 31, 2023

One model for three related jobs: finding a moment from a text query, detecting highlights, and summarizing a video by picking its key clips.

Why it matters: Shows that moment search, highlights and summaries can share one model, which is how newer clipping tools offer 'find the part where...' prompts.

Captions and transcripts

Robust Speech Recognition via Large-Scale Weak Supervision (Whisper)

arXiv, December 6, 2022

OpenAI's Whisper speech recognition model, trained on 680,000 hours of web audio, and released with open weights.

Why it matters: Most auto-caption features and transcript-based clip pickers sit on Whisper or models like it. Caption accuracy sets how much fixing you do before posting.

See the models these papers led to on AI video models.