# VitronEdit: A new unified model for image, video, and 3D editing. Code and weights available — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/f994d09f9900b487

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-27 18:04 ingested 2026-10-28 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xgzw22/vitronedit_a_new_unified_model_for_image_video/)

## retained text



```
I randomly came across this project on Lemmy's Stable Diffusion community and thought it was worth sharing here. Just to be clear, I'm not affiliated with the developers or the project in any way. I simply stumbled upon it, and I couldn't find any discussion about it on Reddit.

 The model is called VitronEdit, and its paper was published on October 19, 2026.

 What caught my attention is that it attempts to bring image editing, video editing, and 3D scene editing into a single unified framework, rather than treating them as completely separate tasks.

 According to the researchers, VitronEdit supports:

 \- Image editing: Modify images using text instructions, optionally with reference images.

\- Video editing: Edit existing videos using text instructions, with optional reference-image conditioning.

\- 3D scene editing: Apply edits across multiple viewpoints while maintaining geometric consistency.

\- Cross-modal learning: Transfer editing knowledge learned from images and videos to 3D scenes.

\- Geometry-aware training: Use depth and camera-pose supervision to improve structural and multi-view consistency.

 One particularly interesting finding in the paper is that applying geometric supervision exclusively to 3D training samples also improved image and video editing performance.

 The project page has visual demonstrations of all three editing modes, including object transformations and multi-view scene editing.

 Links:

 \- Project page and demos: [https://vitron-llm.github.io/VitronEdit/](https://vitron-llm.github.io/VitronEdit/)

\- Research paper: [https://arxiv.org/abs/2610.27049](https://arxiv.org/abs/2610.27049)

\- GitHub (code): [https://github.com/Vitron-LLM/VitronEdit-Code](https://github.com/Vitron-LLM/VitronEdit-Code)

\- Hugging Face (weights): [https://huggingface.co/SanRiiiii/VitronEdit](https://huggingface.co/SanRiiiii/VitronEdit)

\- Original Lemmy post where I discovered it: [https://lemmy.world/post/53470733](https://lemmy.world/post/53470733)

 

The code and model weights are available, although this appears to be a research-oriented release rather than a ready-to-use Stable Diffusion or ComfyUI workflow.

 I haven't tested it myself, so I can't comment on real-world performance, hardware requirements, or how well the demonstrated results generalize.

 Still, the idea of combining image, video, and 3D editing in one framework seems interesting, especially if the community eventually develops easier local inference workflows.

 Has anyone here tried running it locally? I'd be particularly interested in hearing about VRAM requirements, inference speed, and whether anyone is working on ComfyUI integration.

 

Note: I used AI to proofread and correct errors in this post.
```



document fetched 2026-10-28 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xgzw22/vitronedit_a_new_unified_model_for_image_video
