Releases·Tool Release
video-input-to-modern-llm: ffmpeg sampling skill that budgets video frames as tokens for Claude, Codex, any agent
A GitHub repo named video-input-to-modern-llm provides a skill plus two scripts for letting LLMs 'watch' video by sampling frames with ffmpeg instead of ingesting full video. It estimates token cost per model family and fits a selectable token budget by trading frame count against resolution. This matters to practitioners building video-understanding workflows on models without native video input.
105 engagement1 sourceWed, Sep 2, 2026, 08:09 PM
AI Overview
- GitHub repo ships a Claude Code/Codex skill plus scripts that sample video via ffmpeg at chosen resolutions and estimate per-frame token cost before attaching
- Sampling modes include interval, fps, scene-threshold, and keyframes; resolutions cap at 512/768/1024 long side or native and never upscale
- Example run turned 4m12s 1080p video into 15 scene frames at 512x288, estimating ~2955 claude tokens vs ~2160 openai-patch
- Frames get burned-in timestamps and a manifest.; contact sheets tile frames into one image, measured ~1842 claude tokens for 4 sheets
- measure_tokens.py meters attachments, with an --exact mode using Anthropic's counting endpoint; requires only ffmpeg and ffprobe on PATH
Entities
Related
Products·launched
CLI tool extracts video frames and transcript for LLM analysis
A CLI tool called claude-real-video that extracts only changed frames from a video using ffmpeg scene detection and deduplication, aligns them with a timestamped transcript, and provides the folder to any local LLM. It solves the problem that AI models only see transcripts when given video links, missing visual context.
A CLI tool called claude-real-video that extracts only changed frames from a video using ffmpeg scene detection and deduplication, aligns them with a timestamped transcript, and provides the folder to any local LLM. It solves the problem that AI models only see transcripts when given video links, missing visual context.
13 engagement1 sourceredditWed, Jul 15, 2026, 02:57 PM
Social·Community
Put the sampling interval in the prompt when turning a contact sheet into shot JSON
A practitioner demonstrates a reusable prompt for converting video contact sheets into structured shot JSON, using Ling-3.0-flash-VL via OpenRouter. The key technique is stating the frame sampling interval and ordering convention in the prompt so each cell maps to an approximate timestamp. The workflow positions the model output as a first-pass shot inventory to be checked against reference playback before feeding timings into later editing or generation briefs.
A practitioner demonstrates a reusable prompt for converting video contact sheets into structured shot JSON, using Ling-3.0-flash-VL via OpenRouter. The key technique is stating the frame sampling interval and ordering convention in the prompt so each cell maps to an approximate timestamp. The workflow positions the model output as a first-pass shot inventory to be checked against reference playback before feeding timings into later editing or generation briefs.
2 engagement1 sourceredditSat, Oct 3, 2026, 03:21 PM
Social·Community
Community seeks native long-video multimodal models for 6-10 hour streams
A developer building a personal project asks the community how to handle multimodal analysis of very long videos (6-10 hours) locally, noting that current local LLMs with ViT designs suffer quality issues or require frame sampling that defeats native multimodality. The post highlights a gap in tooling for native long-video understanding.
A developer building a personal project asks the community how to handle multimodal analysis of very long videos (6-10 hours) locally, noting that current local LLMs with ViT designs suffer quality issues or require frame sampling that defeats native multimodality. The post highlights a gap in tooling for native long-video understanding.
24 engagement1 sourceredditMon, Aug 10, 2026, 12:06 PM
Developments·Paper
MoD-VLLM: Modularized Dynamic-Granularity Video LLM for Multi-Event Long Video Understanding
Researchers propose MoD-VLLM, a modular Video LLM that dynamically allocates visual token budgets across events in long videos. It addresses the tension between limited tokens and multiple key events by using adaptive capacity allocation and self-correction, improving reliability over two-stage keyframe selection methods.
Researchers propose MoD-VLLM, a modular Video LLM that dynamically allocates visual token budgets across events in long videos. It addresses the tension between limited tokens and multiple key events by using adaptive capacity allocation and self-correction, improving reliability over two-stage keyframe selection methods.
0 engagement1 sourcearxivFri, Jul 17, 2026, 09:21 AM
Releases·Tool Release
Open-source CLI lets Claude watch videos via scene-aware keyframes and timestamped transcripts
A developer released an open-source CLI tool that enables Claude to analyze video content by extracting scene-aware keyframes and timestamped transcripts locally. The tool addresses Claude's inability to process video files directly, allowing users to paste a YouTube URL and get frame-aware answers.
A developer released an open-source CLI tool that enables Claude to analyze video content by extracting scene-aware keyframes and timestamped transcripts locally. The tool addresses Claude's inability to process video files directly, allowing users to paste a YouTube URL and get frame-aware answers.
55 engagement2 sourcesredditThu, Jul 16, 2026, 02:47 AM