VitronEdit open-sources a unified model for image, video, and 3D scene editing
Desperate_Day_5416 · reddit · 2026-10-28
A Reddit user surfaced VitronEdit, a quietly released unified editing model (paper dated Oct 19) that combines image editing, video editing, and 3D scene editing in one framework, with code and weights available.
Capabilities
- Image editing via text instructions, optional reference images
- Video editing of existing footage with optional reference-image conditioning
- 3D scene editing across multiple viewpoints with geometric consistency
- Cross-modal transfer of editing knowledge from images/video to 3D
- Geometry-aware training using depth and camera-pose supervision
Notable finding: applying geometric supervision only to 3D training samples also improved image and video editing performance.
Caveats: the poster hasn't tested it — VRAM needs, inference speed, and generalization are unknown, and it's a research release without ready-made ComfyUI workflows. Links: project page, paper, GitHub, weights on Hugging Face.
More from Multimodal
- Stunning dark etching-style AI illustrations: lurking beasts, starry sails and flower cloaks — OVolosin82152 · 2026-10-28
- Runway launches Ads: autonomous engine that generates and publishes video ad variants — runwayml · 2026-10-28
- Director bot hits Grok Bot Marketplace, turning briefs into finished short films and ads — tetsuoai · 2026-10-28
- Phonon on-device TTS hits 80K downloads with PyTorch-free Rust runtime — mattturck · 2026-10-28
- Halloween set: Batman skateboards with pumpkins, Catwoman pushes a cartful — gizakdag · 2026-10-28
- From vibe-coded games to high-fidelity worlds in months: 'you'll just generate the world' — chrisfirst · 2026-10-28