Hou Xinyu

Ph.D. candidate at MMLab@NTU

[Hou Xinyu @itsmag11]

Most bokeh editors want a sharp photo first. We thought: 𝘄𝗵𝘆 𝗻𝗼𝘁 𝘂𝘀𝗲 𝘁𝗵𝗲 𝗼𝗽𝘁𝗶𝗰𝗮𝗹 𝗰𝗹𝘂𝗲𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗶𝗻 𝘁𝗵𝗲 𝗯𝗹𝘂𝗿? 🔍

Meet AnyBokeh! Edit bokeh directly in blurred photos. No all-in-focus detour. No test-time calibration. ✨

#NeurIPS2026

[Hou Xinyu @itsmag11]

🌐 Project page: https://t.co/BJAHMJeVDY
💻 Inference code: https://t.co/OzFIW0ipwd
🤗 Pretrained models: https://t.co/JzjxM7yXLM

In collaboration with: @csxmli, @OAOAOAOAOAOooo, and @ccloy.

[Hou Xinyu @itsmag11]

Most bokeh editors want a sharp photo first. We thought: 𝘄𝗵𝘆 𝗻𝗼𝘁 𝘂𝘀𝗲 𝘁𝗵𝗲 𝗼𝗽𝘁𝗶𝗰𝗮𝗹 𝗰𝗹𝘂𝗲𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗶𝗻 𝘁𝗵𝗲 𝗯𝗹𝘂𝗿? 🔍

Meet AnyBokeh! Edit bokeh directly in blurred photos. No all-in-focus detour. No test-time calibration. ✨

#NeurIPS2026

[Hou Xinyu @itsmag11]

🎉 Thrilled to share that AnyBokeh has been accepted to NeurIPS 2026!

So happy to receive 5/5/5 from all three reviewers. Huge thanks to my coauthors @csxmli , @OAOAOAOAOAOooo & @ccloy !

🎬 Demo video, project page & GitHub repo coming soon!
#NeurIPS2026

[Hou Xinyu @itsmag11]

Check out the project page to explore the results and try our interactive demo.

[Hou Xinyu @itsmag11]

Last week at #ICCV2025 in Hawaii was a real thrill 🌴 Excited to share our work Omegance, which presents a single tweak in the diffusion denoising process that gives precise control over output granularity.

🎨 Project: https://t.co/Btku0XH9ml
💻 Code: https://t.co/qA6MQvnjl2

[Hou Xinyu @itsmag11]

Thanks for sharing AK!

Come check out our Omegance that allows granularity control in diverse synthesis tasks using one single parameter. Gradio demos and interactive project page are available!

🖤 Code: https://t.co/qA6MQvmLvu
🖤 Project page: https://t.co/yksJ5VVTpY

[AK @_akhaliq]

Omegance

A Single Parameter for Various Granularities in Diffusion-Based Synthesis

[AK @_akhaliq]

Omegance

A Single Parameter for Various Granularities in Diffusion-Based Synthesis

[AK @_akhaliq]

The ArtBench Dataset: Benchmarking Generative Models with Artworks
abs: https://t.co/Zzq0A2i5ob
github: https://t.co/tnHcuprirm

comprises 60,000 images of artwork from 10 distinctive artistic styles, with 5,000 training images and 1,000 testing images per style

[Fei-Fei Li @drfeifei]

Very excited by our work in video prediction to enable robotic planning, which is 512x faster than previous work in autoregressive methods! 🤩 Kudos to the @StanfordSVL team @agrimgupta92 @zhang_yunzhi Stephen Tian, @RobobertoMM @jiajunwu_cs, & thank you @StanfordHAI for support!

[arXiv Daily @Arxiv_Daily]

Video frame interpolation for high dynamic range sequences captured with dual-exposure sensors
https://t.co/H6XwxlsB2j
by Uğur Çoğalan et al.
#Interpolation #NeuralNetwork