Pinned Tweet
๐ [๐ก๐ฒ๐๐ฟ๐๐ฃ๐ฆ ๐ฎ๐ฌ๐ฎ๐ฑ] ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ ๐๐ฐ๐ฐ๐ฒ๐ฝ๐๐ฒ๐ฑ ๐ฎ๐ป๐ฑ ๐๐ผ๐ฑ๐ฒ ๐ฅ๐ฒ๐น๐ฒ๐ฎ๐๐ฒ!
Not everyone is a photo pro โ even with advanced AI tools. Restoring an image to professional grade often requires either ๐ฅ๐ฆ๐ฆ๐ฑ ๐ฑ๐ฉ๐ฐ๐ต๐ฐ๐จ๐ณ๐ข๐ฑ๐ฉ๐บ ๐ฆ๐น๐ฑ๐ฆ๐ณ๐ต๐ช๐ด๐ฆ or ๐ฃ๐ณ๐ฐ๐ข๐ฅ bine many different AI models.
๐ ๐ฆ๐ผ ๐๐ฒ ๐ฎ๐๐ธ๐ฒ๐ฑ ๐ผ๐๐ฟ๐๐ฒ๐น๐๐ฒ๐: Can we build a restoration agent that fully automates this process โ no domain expertise required?
To answer this, we @TAMU partnered with @Topaz and scholars from @Stanford, @Caltech, @UTAustin, @Snap, @ucmerced, @CUBoulder to develop the ๐๐ผ๐ฟ๐น๐ฑโ๐ ๐ณ๐ถ๐ฟ๐๐ ๐ฎ๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฝ๐ต๐ผ๐๐ผ ๐ฟ๐ฒ๐๐๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐๐๐๐ฒ๐บ.
Our system brings together over ๐ฑ๐ฌ ๐๐ฝ๐ฒ๐ฐ๐ถ๐ฎ๐น๐ถ๐๐ฒ๐ฑ ๐๐ ๐บ๐ผ๐ฑ๐ฒ๐น๐ โ for denoising, deblurring, upscaling, face recovery, detail enhancement, and more.
๐ฌ It diagnoses the input image.
๐ฆพ It plans and executes an action graph (just like a human editor).
โ๏ธ After each step, it evaluates the output and adapts the plan if needed.
The results are phenomenal โ as youโll see in the video.
We believe democratizing professional photography will empower anyone to create images suitable for professional use cases. And because we see this as a responsibility, we are ๐ณ๐๐น๐น๐ ๐ผ๐ฝ๐ฒ๐ป-๐๐ผ๐๐ฟ๐ฐ๐ถ๐ป๐ด ๐๐ต๐ฒ ๐๐๐๐๐ฒ๐บ โ to accelerate progress in agentic photo restoration and support the broader community.
๐ ๐๐ต๐ฒ๐ฐ๐ธ ๐ผ๐๐ ๐๐ต๐ฒ ๐ฐ๐ผ๐ฑ๐ฒ: github.com/taco-group/4KAโฆ
๐ ๐ฃ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐ ๐๐ฒ๐ฏ๐ฝ๐ฎ๐ด๐ฒ: 4kagent.github.io
๐'๐บ ๐น๐ผ๐ผ๐ธ๐ถ๐ป๐ด ๐ณ๐ผ๐ฟ๐๐ฎ๐ฟ๐ฑ ๐๐ผ ๐๐ฒ๐ฒ๐ถ๐ป๐ด ๐๐ต๐ฎ๐ ๐๐ผ๐ ๐ฐ๐ฎ๐ป ๐ฏ๐๐ถ๐น๐ฑ ๐๐๐ถ๐ป๐ด ๐ผ๐๐ฟ ๐ผ๐ฝ๐ฒ๐ปโ๐๐ผ๐๐ฟ๐ฐ๐ฒ ๐๐ผ๐ณ๐๐๐ฎ๐ฟ๐ฒ ๐ฎ๐ป๐ฑ ๐๐ผ ๐ฟ๐ฒ๐๐ถ๐ฒ๐๐ถ๐ป๐ด ๐๐ผ๐๐ฟ ๐ฑ๐ฒ๐บ๐ผ ๐ฎ๐ป๐ฑ ๐ณ๐ฒ๐ฒ๐ฑ๐ฏ๐ฎ๐ฐ๐ธ!
Not everyone is a photo pro โ even with advanced AI tools. Restoring an image to professional grade often requires either ๐ฅ๐ฆ๐ฆ๐ฑ ๐ฑ๐ฉ๐ฐ๐ต๐ฐ๐จ๐ณ๐ข๐ฑ๐ฉ๐บ ๐ฆ๐น๐ฑ๐ฆ๐ณ๐ต๐ช๐ด๐ฆ or ๐ฃ๐ณ๐ฐ๐ข๐ฅ bine many different AI models.
๐ ๐ฆ๐ผ ๐๐ฒ ๐ฎ๐๐ธ๐ฒ๐ฑ ๐ผ๐๐ฟ๐๐ฒ๐น๐๐ฒ๐: Can we build a restoration agent that fully automates this process โ no domain expertise required?
To answer this, we @TAMU partnered with @Topaz and scholars from @Stanford, @Caltech, @UTAustin, @Snap, @ucmerced, @CUBoulder to develop the ๐๐ผ๐ฟ๐น๐ฑโ๐ ๐ณ๐ถ๐ฟ๐๐ ๐ฎ๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฝ๐ต๐ผ๐๐ผ ๐ฟ๐ฒ๐๐๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐๐๐๐ฒ๐บ.
Our system brings together over ๐ฑ๐ฌ ๐๐ฝ๐ฒ๐ฐ๐ถ๐ฎ๐น๐ถ๐๐ฒ๐ฑ ๐๐ ๐บ๐ผ๐ฑ๐ฒ๐น๐ โ for denoising, deblurring, upscaling, face recovery, detail enhancement, and more.
๐ฌ It diagnoses the input image.
๐ฆพ It plans and executes an action graph (just like a human editor).
โ๏ธ After each step, it evaluates the output and adapts the plan if needed.
The results are phenomenal โ as youโll see in the video.
We believe democratizing professional photography will empower anyone to create images suitable for professional use cases. And because we see this as a responsibility, we are ๐ณ๐๐น๐น๐ ๐ผ๐ฝ๐ฒ๐ป-๐๐ผ๐๐ฟ๐ฐ๐ถ๐ป๐ด ๐๐ต๐ฒ ๐๐๐๐๐ฒ๐บ โ to accelerate progress in agentic photo restoration and support the broader community.
๐ ๐๐ต๐ฒ๐ฐ๐ธ ๐ผ๐๐ ๐๐ต๐ฒ ๐ฐ๐ผ๐ฑ๐ฒ: github.com/taco-group/4KAโฆ
๐ ๐ฃ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐ ๐๐ฒ๐ฏ๐ฝ๐ฎ๐ด๐ฒ: 4kagent.github.io
๐'๐บ ๐น๐ผ๐ผ๐ธ๐ถ๐ป๐ด ๐ณ๐ผ๐ฟ๐๐ฎ๐ฟ๐ฑ ๐๐ผ ๐๐ฒ๐ฒ๐ถ๐ป๐ด ๐๐ต๐ฎ๐ ๐๐ผ๐ ๐ฐ๐ฎ๐ป ๐ฏ๐๐ถ๐น๐ฑ ๐๐๐ถ๐ป๐ด ๐ผ๐๐ฟ ๐ผ๐ฝ๐ฒ๐ปโ๐๐ผ๐๐ฟ๐ฐ๐ฒ ๐๐ผ๐ณ๐๐๐ฎ๐ฟ๐ฒ ๐ฎ๐ป๐ฑ ๐๐ผ ๐ฟ๐ฒ๐๐ถ๐ฒ๐๐ถ๐ป๐ด ๐๐ผ๐๐ฟ ๐ฑ๐ฒ๐บ๐ผ ๐ฎ๐ป๐ฑ ๐ณ๐ฒ๐ฒ๐ฑ๐ฏ๐ฎ๐ฐ๐ธ!
9
40
378
45,169
Zhengzhong Tu retweeted
ๅฆๆไฝ ๅฌไฟก่ฟไธช๏ผ17ๅฒ้ซไธญๆฏไธๅฐฑ็ตๆดปๅฐฑไธ๏ผๆๅพ
ๅจๅๅ
ซ็บฟๅๅธๆไธชvpnๅฒๆตช็ถๅvibe codingๅ ไธชgithub้กน็ฎ๏ผๅฐฑ่ขซ่ฟๅจๆง้ๅฑฑ็oaiใa\ๆๅฐ็พๅฝๆฅๅทฅไฝ๏ผ้ฃ่ฟ่พๅญๆไบ
่ดนๆผ่ฟๅฅ่ฏๅซ้้่ฟๅจไธๅ๐
โๅฐฝๅฏ่ฝไปฅๆๆ ็บชๅพใ
ๆไธๆฌๅๆๅๅ็ๆนๅผ
ๅชๅๅญฆไน ไฝ ๆๆๅ ด่ถฃ็ๅ ๅฎนโ
22
11
501
68,917
Zhengzhong Tu retweeted
ๅๅ๏ผ็กฎๅฎๆไบใๅพๆฅN้ๅพ่บซ
- ๆๅ็่ฎบ็้กถ็บง่ชๆๅคง่ๅ
- ๆไธๅก่ฟ็ปด็ป้ชๅ
- ๆ่ๆๆบ๏ผsandboxๅคๅนดๆทฑ่ๅ
- ๆ่ฎก็ฎๅญๅจ็ฝ็ปๆฟๅพๅบๆ็้กน็ฎๅ
- ๆๅคฉๆๅฐๅนดๆฉๆฉๅจๆไบๆๆ้กน็ฎๅฎไน ๅนถๆๅ่กจๅ
- ๆๅธๅ ๅธๅงๅฏผๅธไน็ฑป็ๅจ่ฟไบๅฐๆนๅ
- ๆๅ็่ฎบ็้กถ็บง่ชๆๅคง่ๅ
- ๆไธๅก่ฟ็ปด็ป้ชๅ
- ๆ่ๆๆบ๏ผsandboxๅคๅนดๆทฑ่ๅ
- ๆ่ฎก็ฎๅญๅจ็ฝ็ปๆฟๅพๅบๆ็้กน็ฎๅ
- ๆๅคฉๆๅฐๅนดๆฉๆฉๅจๆไบๆๆ้กน็ฎๅฎไน ๅนถๆๅ่กจๅ
- ๆๅธๅ ๅธๅงๅฏผๅธไน็ฑป็ๅจ่ฟไบๅฐๆนๅ
ๅฆๆไฝ ๅฌไฟก่ฟไธช๏ผ17ๅฒ้ซไธญๆฏไธๅฐฑ็ตๆดปๅฐฑไธ๏ผๆๅพ ๅจๅๅ ซ็บฟๅๅธๆไธชvpnๅฒๆตช็ถๅvibe codingๅ ไธชgithub้กน็ฎ๏ผๅฐฑ่ขซ่ฟๅจๆง้ๅฑฑ็oaiใa\ๆๅฐ็พๅฝๆฅๅทฅไฝ๏ผ้ฃ่ฟ่พๅญๆไบ
6
3
76
8,561
ๅฆๆไฝ ๅฌไฟก่ฟไธช๏ผ17ๅฒ้ซไธญๆฏไธๅฐฑ็ตๆดปๅฐฑไธ๏ผๆๅพ
ๅจๅๅ
ซ็บฟๅๅธๆไธชvpnๅฒๆตช็ถๅvibe codingๅ ไธชgithub้กน็ฎ๏ผๅฐฑ่ขซ่ฟๅจๆง้ๅฑฑ็oaiใa\ๆๅฐ็พๅฝๆฅๅทฅไฝ๏ผ้ฃ่ฟ่พๅญๆไบ
่ดนๆผ่ฟๅฅ่ฏๅซ้้่ฟๅจไธๅ๐
โๅฐฝๅฏ่ฝไปฅๆๆ ็บชๅพใ
ๆไธๆฌๅๆๅๅ็ๆนๅผ
ๅชๅๅญฆไน ไฝ ๆๆๅ ด่ถฃ็ๅ ๅฎนโ
22
11
501
68,917
โBrain rotโ was named Oxford Word of the Year in 2024. We wondered: can LLMs get brain rot too? Turns out, yes. ๐ง ๐ซ That question became our #COLM2026 paper, and this work has just been featured by Texas A&M University!
We found that training large language models (LLMs) on low-o โthought skipping,โ causing models to bypass steps in their reasoning and produce incomplete or incorrect answers. The findings underscore an important principle in AI development: Improving the quality of training data is just as critical as improving the models themselves. Read the full story here ๐ tx.ag/BrainRot
@TAMU @UTAustin @NTUsg @LifeAtPurdue @TAMUEngineering
We found that training large language models (LLMs) on low-o โthought skipping,โ causing models to bypass steps in their reasoning and produce incomplete or incorrect answers. The findings underscore an important principle in AI development: Improving the quality of training data is just as critical as improving the models themselves. Read the full story here ๐ tx.ag/BrainRot
@TAMU @UTAustin @NTUsg @LifeAtPurdue @TAMUEngineering
4
1
38
2,997
Machine unlearning erases specific knowledge from a trained foundation model, such as a person's face, private data, or copyrighted content.
We study MLLM unlearning and find 3 evaluation pitfalls that make unlearning look better than it is.
arxiv.org/abs/2610.19613 ๐งต1/
We study MLLM unlearning and find 3 evaluation pitfalls that make unlearning look better than it is.
arxiv.org/abs/2610.19613 ๐งต1/
1
8
41
2,043
The two strongest are among the oldest and simplest: GADiff (2022) and SCRUB (2023). The newest method we test, MANU (2025), ranks 6th and 5th.
Bottom line: no method solves real-world multimodal unlearning yet, and the evaluation bar needs to rise. ๐งต8/
Bottom line: no method solves real-world multimodal unlearning yet, and the evaluation bar needs to rise. ๐งต8/
1
1
1
107
Led by @oscarchew98 Po-Yi Lu, Yi-Jie Cheng, Kunal Jain, @hsuantienlin, and @kuanhaoh_ (@TAMU, @NationalTaiwanU, @Columbia)
๐งต9/
๐งต9/
0
0
1
174
Zhengzhong Tu retweeted
Excited to team up with @googlecloud to put our Live Models in the hands of more creators, developers, and businesses ๐ค
Details: visko.ai/news/visko-andโฆ
Details: visko.ai/news/visko-andโฆ
3
2
6
696
Zhengzhong Tu retweeted
๐ ๐๐ฎ๐ป ๐ณ๐ฟ๐ผ๐ป๐๐ถ๐ฒ๐ฟ ๐ฉ๐๐ ๐ ๐ฑ๐ผ ๐๐ผ๐บ๐ฒ๐๐ต๐ถ๐ป๐ด ๐ฎ๐ ๐๐ถ๐บ๐ฝ๐น๐ฒ ๐ฎ๐ ๐๐๐ฟ๐ป๐ถ๐ป๐ด ๐๐ต๐ฒ๐ถ๐ฟ ๐ต๐ฒ๐ฎ๐ฑ ๐๐ผ ๐ณ๐ถ๐ป๐ฑ ๐ฎ๐ป ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐?
Surprisingly, most still struggle with this basic task.
Excited to share our work, โ๐๐ฒ๐๐ผ๐ป๐ฑ ๐ง๐ต๐ถ๐ป๐ธ๐ถ๐ป๐ด: ๐๐บ๐ฎ๐ป๐ผ๐ถ๐ฑ ๐ฉ๐ถ๐๐๐ฎ๐น ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต,โ accepted to NeurIPS 2026! ๐
Our key finding: ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐บ๐ฎ๐๐๐ฒ๐ฟ๐ for active visual search. ๐ญ๐
Instead of relying on long, multi-turn Chain-of-Thought, we introduce ๐๐ฎ๐ข๐จ๐ช๐ฏ๐ช๐ฏ๐จ ๐ช๐ฏ 360ยฐ:
๐ง ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐๐ต๐ฒ ๐๐ฒ๐บ๐ฎ๐ป๐๐ถ๐ฐ ๐น๐ฎ๐๐ผ๐๐ ๐ฏ๐ฒ๐๐ผ๐ป๐ฑ ๐๐ต๐ฒ ๐ฐ๐๐ฟ๐ฟ๐ฒ๐ป๐ ๐ณ๐ถ๐ฒ๐น๐ฑ ๐ผ๐ณ ๐๐ถ๐ฒ๐
๐งญ ๐๐ฐ๐ ๐ผ๐ป ๐๐ต๐ฒ๐๐ฒ ๐๐ฝ๐ฎ๐๐ถ๐ฎ๐น ๐ฝ๐ฟ๐ถ๐ผ๐ฟ๐ ๐๐ผ ๐ฑ๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฒ๐ฟ๐ฒ ๐๐ผ ๐น๐ผ๐ผ๐ธ ๐ป๐ฒ๐ ๐
This simple idea enables a scalable data engine with ~1.9M training samples ๐ and substantially improves active visual search across both specialized and frontier VLMs.
Perhaps embodied agents donโt always need to think longer โ sometimes, ๐๐ต๐ฒ๐ ๐ป๐ฒ๐ฒ๐ฑ ๐๐ผ ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ. โจ
๐ ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ: arxiv.org/abs/2605.09146
๐ป ๐๐ผ๐ฑ๐ฒ: github.com/Evergreen0929/โฆ
Surprisingly, most still struggle with this basic task.
Excited to share our work, โ๐๐ฒ๐๐ผ๐ป๐ฑ ๐ง๐ต๐ถ๐ป๐ธ๐ถ๐ป๐ด: ๐๐บ๐ฎ๐ป๐ผ๐ถ๐ฑ ๐ฉ๐ถ๐๐๐ฎ๐น ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต,โ accepted to NeurIPS 2026! ๐
Our key finding: ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐บ๐ฎ๐๐๐ฒ๐ฟ๐ for active visual search. ๐ญ๐
Instead of relying on long, multi-turn Chain-of-Thought, we introduce ๐๐ฎ๐ข๐จ๐ช๐ฏ๐ช๐ฏ๐จ ๐ช๐ฏ 360ยฐ:
๐ง ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐๐ต๐ฒ ๐๐ฒ๐บ๐ฎ๐ป๐๐ถ๐ฐ ๐น๐ฎ๐๐ผ๐๐ ๐ฏ๐ฒ๐๐ผ๐ป๐ฑ ๐๐ต๐ฒ ๐ฐ๐๐ฟ๐ฟ๐ฒ๐ป๐ ๐ณ๐ถ๐ฒ๐น๐ฑ ๐ผ๐ณ ๐๐ถ๐ฒ๐
๐งญ ๐๐ฐ๐ ๐ผ๐ป ๐๐ต๐ฒ๐๐ฒ ๐๐ฝ๐ฎ๐๐ถ๐ฎ๐น ๐ฝ๐ฟ๐ถ๐ผ๐ฟ๐ ๐๐ผ ๐ฑ๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฒ๐ฟ๐ฒ ๐๐ผ ๐น๐ผ๐ผ๐ธ ๐ป๐ฒ๐ ๐
This simple idea enables a scalable data engine with ~1.9M training samples ๐ and substantially improves active visual search across both specialized and frontier VLMs.
Perhaps embodied agents donโt always need to think longer โ sometimes, ๐๐ต๐ฒ๐ ๐ป๐ฒ๐ฒ๐ฑ ๐๐ผ ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ. โจ
๐ ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ: arxiv.org/abs/2605.09146
๐ป ๐๐ผ๐ฑ๐ฒ: github.com/Evergreen0929/โฆ
4
7
80
5,233
๐ ๐๐ฎ๐ป ๐ณ๐ฟ๐ผ๐ป๐๐ถ๐ฒ๐ฟ ๐ฉ๐๐ ๐ ๐ฑ๐ผ ๐๐ผ๐บ๐ฒ๐๐ต๐ถ๐ป๐ด ๐ฎ๐ ๐๐ถ๐บ๐ฝ๐น๐ฒ ๐ฎ๐ ๐๐๐ฟ๐ป๐ถ๐ป๐ด ๐๐ต๐ฒ๐ถ๐ฟ ๐ต๐ฒ๐ฎ๐ฑ ๐๐ผ ๐ณ๐ถ๐ป๐ฑ ๐ฎ๐ป ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐?
Surprisingly, most still struggle with this basic task.
Excited to share our work, โ๐๐ฒ๐๐ผ๐ป๐ฑ ๐ง๐ต๐ถ๐ป๐ธ๐ถ๐ป๐ด: ๐๐บ๐ฎ๐ป๐ผ๐ถ๐ฑ ๐ฉ๐ถ๐๐๐ฎ๐น ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต,โ accepted to NeurIPS 2026! ๐
Our key finding: ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐บ๐ฎ๐๐๐ฒ๐ฟ๐ for active visual search. ๐ญ๐
Instead of relying on long, multi-turn Chain-of-Thought, we introduce ๐๐ฎ๐ข๐จ๐ช๐ฏ๐ช๐ฏ๐จ ๐ช๐ฏ 360ยฐ:
๐ง ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐๐ต๐ฒ ๐๐ฒ๐บ๐ฎ๐ป๐๐ถ๐ฐ ๐น๐ฎ๐๐ผ๐๐ ๐ฏ๐ฒ๐๐ผ๐ป๐ฑ ๐๐ต๐ฒ ๐ฐ๐๐ฟ๐ฟ๐ฒ๐ป๐ ๐ณ๐ถ๐ฒ๐น๐ฑ ๐ผ๐ณ ๐๐ถ๐ฒ๐
๐งญ ๐๐ฐ๐ ๐ผ๐ป ๐๐ต๐ฒ๐๐ฒ ๐๐ฝ๐ฎ๐๐ถ๐ฎ๐น ๐ฝ๐ฟ๐ถ๐ผ๐ฟ๐ ๐๐ผ ๐ฑ๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฒ๐ฟ๐ฒ ๐๐ผ ๐น๐ผ๐ผ๐ธ ๐ป๐ฒ๐ ๐
This simple idea enables a scalable data engine with ~1.9M training samples ๐ and substantially improves active visual search across both specialized and frontier VLMs.
Perhaps embodied agents donโt always need to think longer โ sometimes, ๐๐ต๐ฒ๐ ๐ป๐ฒ๐ฒ๐ฑ ๐๐ผ ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ. โจ
๐ ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ: arxiv.org/abs/2605.09146
๐ป ๐๐ผ๐ฑ๐ฒ: github.com/Evergreen0929/โฆ
Surprisingly, most still struggle with this basic task.
Excited to share our work, โ๐๐ฒ๐๐ผ๐ป๐ฑ ๐ง๐ต๐ถ๐ป๐ธ๐ถ๐ป๐ด: ๐๐บ๐ฎ๐ป๐ผ๐ถ๐ฑ ๐ฉ๐ถ๐๐๐ฎ๐น ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต,โ accepted to NeurIPS 2026! ๐
Our key finding: ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐บ๐ฎ๐๐๐ฒ๐ฟ๐ for active visual search. ๐ญ๐
Instead of relying on long, multi-turn Chain-of-Thought, we introduce ๐๐ฎ๐ข๐จ๐ช๐ฏ๐ช๐ฏ๐จ ๐ช๐ฏ 360ยฐ:
๐ง ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐๐ต๐ฒ ๐๐ฒ๐บ๐ฎ๐ป๐๐ถ๐ฐ ๐น๐ฎ๐๐ผ๐๐ ๐ฏ๐ฒ๐๐ผ๐ป๐ฑ ๐๐ต๐ฒ ๐ฐ๐๐ฟ๐ฟ๐ฒ๐ป๐ ๐ณ๐ถ๐ฒ๐น๐ฑ ๐ผ๐ณ ๐๐ถ๐ฒ๐
๐งญ ๐๐ฐ๐ ๐ผ๐ป ๐๐ต๐ฒ๐๐ฒ ๐๐ฝ๐ฎ๐๐ถ๐ฎ๐น ๐ฝ๐ฟ๐ถ๐ผ๐ฟ๐ ๐๐ผ ๐ฑ๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฒ๐ฟ๐ฒ ๐๐ผ ๐น๐ผ๐ผ๐ธ ๐ป๐ฒ๐ ๐
This simple idea enables a scalable data engine with ~1.9M training samples ๐ and substantially improves active visual search across both specialized and frontier VLMs.
Perhaps embodied agents donโt always need to think longer โ sometimes, ๐๐ต๐ฒ๐ ๐ป๐ฒ๐ฒ๐ฑ ๐๐ผ ๐ถ๐บ๐ฎ๐ด๐ถ๐ป๐ฒ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ. โจ
๐ ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ: arxiv.org/abs/2605.09146
๐ป ๐๐ผ๐ฑ๐ฒ: github.com/Evergreen0929/โฆ
4
7
80
5,233
Zhengzhong Tu retweeted
Weโre getting together with the Google Startup team at SF Tech Week to talk about Viskoโs Live Models and Orbis, our foundation model for real-time interactive world generation.
Weโll explore why TPUs are particularly well suited for the demanding inference workloads behind Live es to bring these real-time applications to scale.
Save your spot! RSVP link in the comments โก๏ธ
Weโll explore why TPUs are particularly well suited for the demanding inference workloads behind Live es to bring these real-time applications to scale.
Save your spot! RSVP link in the comments โก๏ธ
4
3
15
6,097
Zhengzhong Tu retweeted
Can frontier VLMs do something as simple as turning their head to find an object?
Turns out, most struggle at this basic 2-DoF task.
While we discover that, Imagnation in 360ยฐ matters in active visual search!๐คฏ
๐ Accepted to NeurIPS 2026! ๐
1/n
Turns out, most struggle at this basic 2-DoF task.
While we discover that, Imagnation in 360ยฐ matters in active visual search!๐คฏ
๐ Accepted to NeurIPS 2026! ๐
1/n
9
3
9
864
Excited to share our new work, ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฒ๐ป๐ฐ๐ต: ๐๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ๐ถ๐ป๐ด ๐๐ถ๐ด๐ต-๐๐ถ๐ฑ๐ฒ๐น๐ถ๐๐ ๐ฉ๐ถ๐ฑ๐ฒ๐ผ ๐ฆ๐ฐ๐ฒ๐ป๐ฒ ๐ง๐ฒ๐
๐ ๐๐ฑ๐ถ๐๐ถ๐ป๐ด, accepted to the #NeurIPS 2026 ๐&๐ ๐ง๐ฟ๐ฎ๐ฐ๐ธ! ๐
Editing text in videos is surprisingly challenging: a good model needs to genekeep it temporally consistent across frames, and preserve everything else in the scene.
To study this problem better, we introduce ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฒ๐ป๐ฐ๐ต, a comprehensive benchmark for video scene text editing. Our release includes:
๐น ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฎ๐๐ฎ๐๐ฒ๐ โ 387 real-world 720p videos with text-region masks and editing instructions
๐น ๐ ๐๐ต๐ฟ๐ฒ๐ฒ-๐ฎ๐ ๐ถ๐ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ฝ๐ฟ๐ผ๐๐ผ๐ฐ๐ผ๐น covering text correctness, temporal quality, and edit locality with 13 metrics
๐น ๐๐ ๐๐ฒ๐ป๐๐ถ๐๐ฒ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ผ๐ณ ๐ด ๐ฏ๐ฎ๐๐ฒ๐น๐ถ๐ป๐ฒ๐ across four video-editing paradigms
๐น ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฑ๐ถ๐-๐ญ๐ฐ๐, an open-source reference model with motion-aligned glyph-video conditioning
๐น Open-source ๐ฑ๐ฎ๐๐ฎ, ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ฐ๐ผ๐ฑ๐ฒ, ๐บ๐ผ๐ฑ๐ฒ๐น ๐๐ฒ๐ถ๐ด๐ต๐๐, ๐ฎ๐ป๐ฑ ๐น๐ฒ๐ฎ๐ฑ๐ฒ๐ฟ๐ฏ๐ผ๐ฎ๐ฟ๐ฑ
One takeaway from our study: no single metric tells the whole story. Models that preserve the scene well may fail to actually edit the text, while models that generate the right text frame-by-frame can still suffer from severe temporal flickering.
ViTeX-Bench is designed to make these trade-offs explicit and reproducible. We hope ViTeX-Bench can provide a useful foundation for future research on precise and temporally consistent video editing.
๐ Project page: vitex-bench.github.io
๐ Paper link: arxiv.org/abs/2609.40356
#NeurIPS2026 #ComputerVision #GenerativeAI #VideoEditing #VideoGeneration #MultimodalAI #MachineLearning #AIResearch
Editing text in videos is surprisingly challenging: a good model needs to genekeep it temporally consistent across frames, and preserve everything else in the scene.
To study this problem better, we introduce ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฒ๐ป๐ฐ๐ต, a comprehensive benchmark for video scene text editing. Our release includes:
๐น ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฎ๐๐ฎ๐๐ฒ๐ โ 387 real-world 720p videos with text-region masks and editing instructions
๐น ๐ ๐๐ต๐ฟ๐ฒ๐ฒ-๐ฎ๐ ๐ถ๐ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ฝ๐ฟ๐ผ๐๐ผ๐ฐ๐ผ๐น covering text correctness, temporal quality, and edit locality with 13 metrics
๐น ๐๐ ๐๐ฒ๐ป๐๐ถ๐๐ฒ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ผ๐ณ ๐ด ๐ฏ๐ฎ๐๐ฒ๐น๐ถ๐ป๐ฒ๐ across four video-editing paradigms
๐น ๐ฉ๐ถ๐ง๐ฒ๐ซ-๐๐ฑ๐ถ๐-๐ญ๐ฐ๐, an open-source reference model with motion-aligned glyph-video conditioning
๐น Open-source ๐ฑ๐ฎ๐๐ฎ, ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป ๐ฐ๐ผ๐ฑ๐ฒ, ๐บ๐ผ๐ฑ๐ฒ๐น ๐๐ฒ๐ถ๐ด๐ต๐๐, ๐ฎ๐ป๐ฑ ๐น๐ฒ๐ฎ๐ฑ๐ฒ๐ฟ๐ฏ๐ผ๐ฎ๐ฟ๐ฑ
One takeaway from our study: no single metric tells the whole story. Models that preserve the scene well may fail to actually edit the text, while models that generate the right text frame-by-frame can still suffer from severe temporal flickering.
ViTeX-Bench is designed to make these trade-offs explicit and reproducible. We hope ViTeX-Bench can provide a useful foundation for future research on precise and temporally consistent video editing.
๐ Project page: vitex-bench.github.io
๐ Paper link: arxiv.org/abs/2609.40356
#NeurIPS2026 #ComputerVision #GenerativeAI #VideoEditing #VideoGeneration #MultimodalAI #MachineLearning #AIResearch
0
4
35
2,620