
Senior Expert at Bosch Center for Artificial Intelligence (BCAI) @Bosch_AI, leading a Bosch Industry on Campus (IoC) group at University of Tübingen
Baden-Württemberg, Germany
Joined January 2021
- Tweets61
- Following156
- Followers471
- Likes315
Pinned Tweet
Excited and honored to receive Robert Bosch Inventor Award 2024 alongside my dear colleague @anna_khoreva .
This achievement wouldn’t have been possible without the support of everyone at @Bosch_AI .
Grateful for this recognition and looking forward to more innovation together!
Dan Zhang retweeted
Call for Papers: 2nd Workshop on Test-Time Scaling for Computer Vision (ViSCALE) at CVPR 2026! @CVPR
📢 Focus this year: test-time scaling in vision, spatial reasoning, world models, embodied intelligence, and safety.
📅 Deadline: March 10, 2026
🔗 viscale.github.io
Dan Zhang retweeted
NEWS: Chinese media tested ADAS in various scenarios, including highways & night driving. @Tesla’s vision-based system outperformed emerging Chinese brands like Huawei & Xiaomi, as well as traditional automakers. Even with LiDAR, competitors’ ADAS performance lags behind Tesla.
Rankings
Top Performers:
• Tesla Model 3 (RWD Refreshed) – 5/6
• Tesla Model X (2023 LR Refreshed) – 5/6
• Xpeng G6 – 3/6
• Wenjie M9 – 3/6
• Zhijle R7 – 3/6
• BYD Z9GT EV – 3/6
Mid-Tier:
• Aion RT (2025) – 2/6
• Platinum 3X – 2/6
• Avita 12 – 2/6
• Wenjie M7 – 2/5
• Avita 07 – 2/5
Low Performers:
• Ideal L6 – 1/6
• Xiaomi SU7 – 1/5
• Wenjie M8 – 1/5
• QinLDM – 1/5
• iCAR V23 – 1/5
• Xiaomi SU7 Ultra (2025) – 1/4
• BYD Seagull – 1/4
• NIO ES6 – 1/4
Failed All Tests (0 Passes):
• Zeekr 001 (2025 YOU Edition) – 0/6
• Baojun Enjoy PHEV – 0/6
• Lynk & Co – 0/6
• Han LEV – 0/5
• Zero runing C10 – 0/5
• PASSAT – 0/5
• GAC Honda P7 – 0/5
• Zeekr 7X (2025 100kWh) – 0/5
• Xpeng P7+ – 0/4
• Song Pro DM – 0/4
• Letao L60 – 0/4
• Star Era ET – 0/4
• Firefly – 0/4
Full 1.5 hour video in thread below:
👇checkout our work that greatly improves text annotations of segmentation and object detection datasets. #NeurIPS2024
🔥 Are you ever dissatisfied with the imprecise names in vision-language datasets?
🚀 At #NeurIPS2024, we introduce 𝐑𝐄𝐍𝐎𝐕𝐀𝐓𝐄, showing how better segmentation dataset names lead to 𝐛𝐞𝐭𝐭𝐞𝐫 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 & 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧.
Let’s dive in! 🧵👇
Dan Zhang retweeted
🔥 Are you ever dissatisfied with the imprecise names in vision-language datasets?
🚀 At #NeurIPS2024, we introduce 𝐑𝐄𝐍𝐎𝐕𝐀𝐓𝐄, showing how better segmentation dataset names lead to 𝐛𝐞𝐭𝐭𝐞𝐫 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 & 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧.
Let’s dive in! 🧵👇
Dan Zhang retweeted
If this caught your attention, have a look at out paper and come talk to us during the NeurIPS poster session on Thu 12 Dec at 11:00 a.m. PST (East Exhibit Hall A-C #4505)
Great collaboration with: @AlexTimans, @eric_nalisnick, @canaesseth, @isDanZhang, @HvTin and others
Dan Zhang retweeted
🤗Code is out: github.com/boschresearch/VST…
🚀TL;DR: VSTAR generates longer videos with dynamic visual evolution in a single pass. No fine-tuning is needed!
🙌Check out our project page: yumengli007.github.io/VSTAR/
Bill Beluch @margret_keuper @isDanZhang @anna_khoreva @Bosch_AI ❤️
Thanks @_akhaliq for sharing! VSTAR enables long video synthesis with dynamic visual evolution in a single pass! Using a pretrained T2V model, no finetuning is required! We also analyze temporal attention across models and provide valuable insights.
yumengli007.github.io/VSTAR
Things change all the time, and we cannot stop that. Check out our poster if you don’t want to keep retraining your deployed models!
Drop by our poster at the #ICML2024 SPIGM Workshop on Friday!
With @eric_nalisnick and @isDanZhang we propose probabilistic state-space models to adapt a classifier to distribution shift without seeing any labels.
📍SPIGM Workshop, Lehar 3, 15:10-16:10
Dan Zhang retweeted
Check out our recent work on parameter efficient finetuning of LLMs and VLMs - ETHER, which was accepted at @icmlconf ! #ICML2024
github.com/mwbini/ether
More info 👇
Dan Zhang retweeted
VSTAR
Generative Temporal Nursing for Longer Dynamic Video Synthesis
Despite tremendous progress in the field of text-to-video (T2V) synthesis, open-sourced T2V diffusion models struggle to generate longer videos with dynamically varying and evolving content. They tend to
Dan Zhang retweeted
Check out our new work on Generative Temporal Nursing for Longer Dynamic Video Synthesis 👇
yumengli007.github.io/VSTAR/
Thanks @_akhaliq for sharing!
Check out our recent work on improving video synthesis, very little extra effort for generating longer and much more dynamic videos using pretrained T2V models.
Thanks @_akhaliq for sharing! VSTAR enables long video synthesis with dynamic visual evolution in a single pass! Using a pretrained T2V model, no finetuning is required! We also analyze temporal attention across models and provide valuable insights.
yumengli007.github.io/VSTAR
Dan Zhang retweeted
Thanks @RSIPvision for featuring our #ICLR2024 paper "Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive"!
Check out our project page with code & models 🤗 yumengli007.github.io/ALDM
Joint work with @margret_keuper @isDanZhang @anna_khoreva at @Bosch_AI
Here is Computer Vision News of March 2024.
New: first lesson in robotics by awesome @madibabaiasl Madi Babaiasl.
rsipvision.com/ComputerVisio…
Enjoy the reading and - if you like it - share with friends!
It's a publication by @RSIPvision.
Subscribe for free on page 34.
👇Checkout our paper just accepted at ICLR2024. Adversarial supervision w. multi-denoising step unrolling makes Layout-to-Image diffusion models thrive. 🥳
🚀ALDM is accepted at #ICLR2024 @iclr_conf !
We make layout-to-image diffusion models thrive, by integrating Adversarial Supervision to provide explicit alignment feedback & Multistep Unrolling for consistent adherence!
Code & Models are released🤗 yumengli007.github.io/ALDM
Dan Zhang retweeted
Check out our #NeurIPS2023 poster at the Workshop on Distribution Shifts on Friday (Room R06-R09)!
The work with @isDanZhang and @eric_nalisnick explores divergence-based disagreement notions for predicting performance under distribution shift.
Paper: arxiv.org/abs/2312.08033
Our models struggle at generating high-quality uncertainty estimates, because they are not trained for doing that. Our paper at #NeurIPS2023 provided a simple solution.
Visit us at the poster session #502 Great hall & Hall B1+B2 (Wed 13, 7:15-9.15 a.m. CST).
Dan Zhang retweeted
Tired of overthinking decisions in your life? I can't help you with that, but come to our NeurIPS poster #514 tomorrow (Tue, 12th Dec) between 5-7pm to see how to prevent your early-exit neural networks from overthinking
👇Check out our work presented at NeurIPs now!
Early-exit networks aim at adjusting the computation effort according to the actual complexity of classifying each sample. We further encourage them to become gradually confident. More processing only for better prediction quality!
Replying to @MetodJazbec
This was an awesome collaboration with
@JamesAllingham, @isDanZhang and my supervisor @eric_nalisnick
If you are at BMVC2023, please meet us at our oral talk tomorrow at 16:30 (UK GMT time). We will present a simple solution to help your T2I model better follow your instructions.
🚀 Thrilled to announce that I'll be presenting our Oral paper "Divide & Bind" at #BMVC2023 tomorrow (21.11) at 16:30 (UK GMT time) 🗓️!
What to expect: Witness incredible text-to-image synthesis powered by improved Stable Diffusion, and the best part? No fine-tuning required! 🌟
1
481


























