Pingchuan Ma

Thanks, my dear investor!

Huge thanks to all my great collaborators, @mgui7.bsky.social, @joh-schb.bsky.social , Xiaopei Yang, Yusong Li, Felix Krause, Olga Grebenkova, @vtaohu.bsky.social, and Björn Ommer at @compvis.bsky.social.

I am excited to connect or catch up in person at #ICCV2025. I am also seeking full-time or internship opportunities 🧑🏻‍💻 starting June 2026.

2️⃣ ArtFM: Stochastic Interpolants for Revealing Stylistic Flows across the History of Art 🎨
TL;DR: Modeling how artistic style evolves over 500 years without relying on ground-truth pairs.
📍 Poster Session 2
🗓️ Tue, Oct 21 — 3:00 PM
🧾 Poster #80
🔗 compvis.github.io/Art-fm/

1️⃣ SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models 🔄
TL;DR: We learn disentangled representations implicitly by training only on merging them.
📍 Poster Session 4
🗓️ Wed, Oct 22 — 2:30 PM
🧾 Poster #3
🔗 compvis.github.io/SCFlow/

I’m thrilled to share that I’ll present two first-authored papers at #ICCV2025 🌺 in Honolulu together with @mgui7.bsky.social ! 🏝️
(Thread 🧵👇)

🤔 What happens when you poke a scene — and your model has to predict how the world moves in response?

We built the Flow Poke Transformer (FPT) to model multi-modal scene dynamics from sparse interactions.

It learns to predict the 𝘥𝘪𝘴𝘵𝘳𝘪𝘣𝘶𝘵𝘪𝘰𝘯 of motion itself 🧵👇

Jokes aside, I understand the submission load is massive, but shifting the burden of filtering invalid or duplicate entries onto reviewers isn’t a good incentive. It is 2025, and simple automated checks should be easy to employ, help the process, and respect reviewers’ time.

I just wrapped up my bidding process for AAAI 26, which is always an enjoyable experience. This year, I came across submissions titled things like “000ASD”, “testy”, “123321241”, and even cases with identical titles and abstracts but different submission numbers.

this sounded way to intimate

Our work received an invited talk at the Imageomics-AAAI-25 workshop of #AAAI25. @vtaohu.bsky.social will be representing us there. Without me being there, I still would like to share our poster with you :D

We also have another oral presentation for DepthFM on March 1, 2:30 pm-3:45 pm.

(reference source in the Alt text of the figures due to character limit)

Interestingly, LLM-assigned descriptions behave similarly indistinctively as randomly assigned descriptions.

We also achieve competitive results in the conventional eval setting.

🚀 We introduced a training-free approach to select small yet meaningful neighborhoods of discriminative descriptions that improve classification accuracy. We validated our findings across seven datasets, showing consistent gains without distorting the underlying VLM embedding space.

Interestingly, LLM-assigned descriptions behave similarly indistinctively as randomly assigned descriptions in this case.

🤔 To check out what happened, we proposed a new evaluation scenario to isolate the semantic impact by eliminating noise augmentation, ensuring only meaningful descriptions are useful. Random strings and unrelated texts degrade performance, while our assigned descriptions show clear advantages.

random string or unrelated text can have a similar effect, source: arxiv.org/abs/2306.07282

🤯 However, another recent study showed that descriptions augmented by unrelated text or noise could achieve similar performances. This raises the question: Does genuine semantics drive this gain, or is it an ensembling effect (similar to Multi-Crop test-time augmentation in the vision case)?

prompting LLMs for additional descriptions, source: arxiv.org/abs/2210.07183

Studies have shown that VLMs performance can be boosted by prompting LLMs for more detailed descriptions. Then use these descriptions to augment the original classname during inference, such as “tiger, a big cat; with sharp teeth” instead of simply “tiger”.

This work was co-led by x.com/LennartRietdorf, and collaborated with Dmytro Kotovenko, @vtaohu.bsky.social, and Björn Ommer.

Our method pipeline

🤔When combining Vision-language models (VLMs) with Large language models (LLMs), do VLMs benefit from additional genuine semantics or artificial augmentations of the text for downstream tasks?

🤨Interested? Check out our latest work at #AAAI25:

💻Code and 📝Paper at: github.com/CompVis/DisCLIP

🧵👇

🤔 Why do we extract diffusion features from noisy images? Isn’t that destroying information?

Yes, it is - but we found a way to do better. 🚀

Here’s how we unlock better features, no noise, no hassle.

📝 Project Page: compvis.github.io/cleandift
💻 Code: github.com/CompVis/clea...

🧵👇