xiangpeng yang

a PhD student at UTS

[xiangpeng yang @Ayden_Yang_]

📢 Excited to share that our unified video editing work VideoCoF is accepted to CVPR 2026！🚀
🎬TL;DR: We propose a Chain-of-Frames framework for unified video editing with 16× length extrapolation (512+ frames).
code: https://t.co/gAg8rqrwaK
#CVPR2026 #ECCV2026 #Seedance

[xiangpeng yang @Ayden_Yang_]

Thanks HuggingFace for sharing! 2-30s per video with only 4-step infer, try it!

[DailyPapers @HuggingPapers]

Dive into the details and try it out! VideoCoF performs a "see, reason, then edit" procedure, enabling fine-grained control and motion alignment.

Paper: https://t.co/K241LqJyJO
Model: https://t.co/7jejipg13I

[xiangpeng yang @Ayden_Yang_]

@HuggingPapers @karpathy @jimmy_dore @huggingface @arankomatsuzaki @fofrAI @LiorOnAI https://t.co/wE8inQ3KvG already submitted it~ just for better impact and PR.

[xiangpeng yang @Ayden_Yang_]

Can video diffusion models reason over spatial content? 🤔

We propose VideoCoF, a unified framework reasoning video content via Chain of Frames.

🧠 Seeing, Reasoning, then Editing
📈 Infinite Length generalization
🎯 spatial aware, instance & part-level editing

[xiangpeng yang @Ayden_Yang_]

📄 Paper: https://t.co/Ez3CvAciHR
🏠 Project Page: https://t.co/Jv1yvjQyHC
💻Code: https://t.co/gAg8rqrwaK

[DailyPapers @HuggingPapers]

Unified Video Editing with Temporal Reasoner (VideoCoF)

This novel Chain-of-Frames approach enables precise, mask-free video editing and 4x length extrapolation, achieving SOTA performance with just 50k training pairs.

[xiangpeng yang @Ayden_Yang_]

🎉Our pioneering work VideoGrain (#ICLR2025) can also localize concept and editing, check our demo~
#VideoGeneration #Diffusion #CVPR2025 #ArtificialIntelligence #ML @_akhaliq @Gradio

Different from simple localized concepts #conceptattention @alec_helbling , we can localize multi-concepts at one denoising time, which means we can edit multi-regions with just inference once time!

👇check how #VideoGrain localizes multi-concepts and edit at once:

[xiangpeng yang @Ayden_Yang_]

🎉Our pioneering work VideoGrain (#ICLR2025) can also localize concept and editing, check our demo~

Different from simple localized concepts, we can localize multi-concepts at one denoising time, which means we can edit multi-regions with just inference once time!

👇check how #VideoGrain localizes multi-concepts and edit at once:

[Alec Helbling @alec_helbling]

Create heatmaps that localize text concepts in generated videos.

We discovered that our approach, ConceptAttention, can be directly extended from image generation to video generation models!

It's amazing how simple techniques often generalize way better than more complex ones.

[xiangpeng yang @Ayden_Yang_]

#VideoGrain #ICLR2025 #VideoGeneration #artificalintelligence

📷 Details (3/n) - Method

So our key insights and designs are:

(1) Text-to-region control. As seen in the figure mid, we modulate cross-attn and redistribute each local prompt’s weight to align with target areas, enabling precise text-to-region control.

(2) Keep feature separation. As seen in the figure bottom, we modulate self-attn, the query feature from the left man’s nose (e.g., p) attends only to the left instance, avoiding distraction to the right instance.
We break the diffusion model’s class-level feature correspondence, ensuring feature separation at the instance level.

[xiangpeng yang @Ayden_Yang_]

🚀 #ICLR2025 Thanks @_akhaliq @Gradio share.
We just open-sourced VideoGrain, a training-free, multi-grained (class/instance/part) video editing method

Check it out:

📄ArXiv: https://t.co/q700znUrCh

🎥Project: https://t.co/5vIvJmVn7Y

💻Code: https://t.co/bWfbH6hrjs

[xiangpeng yang @Ayden_Yang_]

#VideoGrain #ICLR2025 #VideoGeneration #ArtificialIntelligence
📷 Details (2/n) - Motivation

So why diffusion model fail at multi-grained video editing?

1) As seen in (b), the clustering inversion self-attn feature shows that although it captures a clear semantic layout, it fails to distinguish between distinct instances. (e.g., “left man” and “right man”).

2) Direct editing same-class instances fails since the cross-attn weight distribution is not right, as shown in (d). i.e., "Iron Man" and "Spiderman" weights overlap on the left, and “Blossom” weight leaks onto the right man, resulting in the failed edit in (c).

[xiangpeng yang @Ayden_Yang_]

🚀 #ICLR2025 Thanks @_akhaliq @Gradio share.
We just open-sourced VideoGrain, a training-free, multi-grained (class/instance/part) video editing method

Check it out:

📄ArXiv: https://t.co/q700znUrCh

🎥Project: https://t.co/5vIvJmVn7Y

💻Code: https://t.co/bWfbH6hrjs

[xiangpeng yang @Ayden_Yang_]

🚀 #ICLR2025 Thanks @_akhaliq @Gradio share.
We just open-sourced VideoGrain, a training-free, multi-grained (class/instance/part) video editing method

Check it out:

📄ArXiv: https://t.co/q700znUrCh

🎥Project: https://t.co/5vIvJmVn7Y

💻Code: https://t.co/bWfbH6hrjs

[xiangpeng yang @Ayden_Yang_]

#VideoGrain
📷 Details (1/n)👇

What is multi-grained video editing?
As seen in the fig left:

(1) Class-level: Editing objects within the same class.

(2) Instance-level: Editing each individual instance to distinct object.

(3) Part-level: Adding new objects or modifying existing attributes at the part-level

However, as seen in the fig right, previous SOTA failed at the instance level, like Pika.

[xiangpeng yang @Ayden_Yang_]

🚀 #ICLR2025 Thanks @_akhaliq @Gradio share.
We just open-sourced VideoGrain, a training-free, multi-grained (class/instance/part) video editing method

Check it out:

📄ArXiv: https://t.co/q700znUrCh

🎥Project: https://t.co/5vIvJmVn7Y

💻Code: https://t.co/bWfbH6hrjs

[xiangpeng yang @Ayden_Yang_]

@ED84VG @_akhaliq DALL·E is a closed-source image model, while our VideoGrain is fully open-sourced and designed for video. It’s training-free and accessible, letting anyone build or adapt it freely.

[xiangpeng yang @Ayden_Yang_]

@UniversaAI @_akhaliq Thanks~🥰

[xiangpeng yang @Ayden_Yang_]

🚀 #ICLR2025 Thanks @_akhaliq @Gradio share.
We just open-sourced VideoGrain, a training-free, multi-grained (class/instance/part) video editing method

Check it out:

📄ArXiv: https://t.co/q700znUrCh

🎥Project: https://t.co/5vIvJmVn7Y

💻Code: https://t.co/bWfbH6hrjs

[xiangpeng yang @Ayden_Yang_]

@UniversaAI @_akhaliq thanks! check our latest code: https://t.co/bWfbH6hrjs