Stanislav Frolov

I am happy to share that SpotDiffusion was accepted to WACV 2025. Page: spotdiffusion.github.io Code: github.com/stanifrolov/sp… Paper: arxiv.org/abs/2407.15507 SpotDiffusion is an efficient method for seamless panorama generation from text. 🧵

Current image generation models produce stunning images, but do they perform well as synthetic training data generators? In our new CVPR paper, we observe a surprising performance regression and investigate why. Link to paper: arxiv.org/abs/2602.19946 Congrats to all co-authors!

Happy to share our VXAI paper at TMLR, a review & framework for the evaluation of XAI. Check out the VXAI explorer here vxai.dfki.de

Unifying VXAI: A Systematic Review and Framework for the Evaluation of Explainable AI David Dembinsky, Adriano Lucieri, Stanislav Frolov, Hiba Najjar, Ko Watanabe, Andreas Dengel. Action editor: Krikamol Muandet. openreview.net/forum?id=wAvFL… #explanation

Say hello to DINOv3 🦖🦖🦖 A major release that raises the bar of self-supervised vision foundation models. With stunning high-resolution dense features, it’s a game-changer for vision tasks! We scaled model size and training data, but here's what makes it special 👇

Happy to share that TKG-DM, a training-free chroma key content generation diffusion model was accepted to CVPR 25. Project led by @Oguryu417 Paper: arxiv.org/abs/2411.15580 Code: github.com/ryugo417/TKG-DM

Checkout PromptMap, presented at IUI'25, a new interaction style with text-to-image models/data that allows users to freely explore a vast collection of synthetic prompts through a map-like view with semantic zoom. Paper: arxiv.org/abs/2503.09436 Code: github.com/Bill2462/promp…

@vikhyatk Yes and it can have huge impact when you evaluate quality of generative models on images, for example with FID.

if you train a model exclusively on JPEG images, will performance drop on other image file formats?

I received feedback that my post about reviews not being "random" caused stress for some students. I'm sorry for that. It was meant to be empowering. Personally, I find the idea that I don't have some control over the destiny of my papers to be disheartening. If the process is random and I have no influence, why would I bother? Taking personal responsibility for poor reviews is, for me, a way to take control. It gives me a chance to act. Even better, it gives me hugely valuable feedback that my work missed the mark. Mea culpa. Saying the reviewers are bad or reviews are random may give temporary solace -- "it's someone else's fault" -- but it doesn't lead to long-term success and, ultimately, satisfaction. If you found my post disheartening, then see my guide for how to write a good CVPR paper. It's an action plan for writing a paper that reviewers will understand. It's not easy to do everything I describe and it takes practice. But a PhD is about practice. We practice the whole process of doing science and a big part of this is practicing writing about our science. I will also share my tips for writing rebuttals once I find time to clean them up. perceiving-systems.blog/post/writing-a…

I am very happy to share that our open-access survey about diffusion models in the field of image super-resolution got accepted by #IEEE TNNLS: ieeexplore.ieee.org/document/10737… #images #diffusion #models #survey @stfrolov @rave78 @spalaciob

We propose a time-dependent, attention-guided masking approach that prioritizes high-attention regions first, gradually refining the entire image. This improves quality across various models. Paper: arxiv.org/abs/2308.07977 Thanks to @LuckyOwl95 @rave78 @spalaciob @DFKI

We find that important image pixels, as measured by the attention values of DINO, are more challenging to learn (higher reconstruction error).

Dynamic attention-guided diffusion accepted to #WACV2025 🎉 We challenge the common SR diffusion approach: must the entire image be updated at each step? Some regions, like faces, may need more focus than plain backgrounds. 🧵

Check out our project page and paper for more visual results. Page: spotdiffusion.github.io Code: github.com/stanifrolov/sp… Paper: arxiv.org/abs/2407.15507 Thanks to my collaborators @bmoser95 and Andreas Dengel. @DFKI @rptu_kl_ld @wacv_official

We can produce seamless panoramas much faster by leveraging the iterative nature of diffusion models and shifting non-overlapping denoising windows over time.

I am happy to share that SpotDiffusion was accepted to WACV 2025. Page: spotdiffusion.github.io Code: github.com/stanifrolov/sp… Paper: arxiv.org/abs/2407.15507 SpotDiffusion is an efficient method for seamless panorama generation from text. 🧵

✨🎨🏰Super excited to share our new paper Ensemble everything everywhere: Multi-scale aggregation for adversarial robustness Inspired by biology we 1) get adversarial robustness + interpretability for free, 2) turn classifiers into generators & 3) design attacks on vLLMs 1/12

I can’t find a recent paper (and tweet) that had emojis all over an image. I think it was a method about interpreting (possibly segmenting) images with/from diffusion models. Can somebody help?

Wow that’s cool! LoRA but for training.

Thanks @_akhaliq for promoting our work! With GaLore, now it is possible to pre-train a 7B model in NVidia RTX 4090s with 24G memory! 🤔How? Instead of assuming low-rank weight structure like LoRA, we show that the weight gradient is naturally low-rank and thus can be projected into a (changing) low-dimensional space. Therefore, we save memory on gradient, Adams' momentum and variance at the same time! As a result, unlike LoRA, GaLore does not change the training dynamics and can be used to pre-train a 7B model from scratch, without any memory-consuming warm-up. This yields 1B/7B models with comparable perplexity as vanilla training up to 13B/20B tokens, using only 1/4 of the rank. With 1/2 of the rank, our 1B model is even better🤯. GaLore can also be used to do fine-tuning as well, yielding comparable results with LoRA. Thanks to awesome collaborators @jiawzhao, @KyriectionZhang, @BeidiChen, Zhangyang Wang and @AnimaAnandkumar!