Nikita Durasov

VLMs & Agents @NVIDIAAI, PhD @EPFL_en, making machines see and reason, prev @Apple, @Amazon, @SamsungResearch

[Nikita Durasov @norange_ai]

1/4🎉 Our paper "𝐃𝟑𝐃𝐑: Diffusion Models are Secretly Zero-Shot 3DGS Harmonizers" is in TMLR!

Diffusion models implicitly learn scene lighting — we leverage this for zero-shot 3DGS object insertion & relighting. No env maps. No material estimation.

🌐 https://t.co/wRwUbQDtND

[Nikita Durasov @norange_ai]

@giffmana Oh, nice, I almost can see my house on the right side over the lake!

[Nikita Durasov @norange_ai]

Dive into our work below:
📄 Paper: https://t.co/MbvHQh9xBi
📝 arXiv: https://t.co/ah34XptjEK
💻 Code: https://t.co/m0fhYRgE62🌐 Project: https://t.co/yEeoqQAUnM
🎬 Video: https://t.co/phibJVpnWp

#CVPR2026 #AdvML #VisionLanguageModels #AdversarialDefense

[Nikita Durasov @norange_ai]

1/6 🏆 Thrilled our paper, "MirrorCheck: Efficient Adversarial Defense for Vision-Language Models," won the Distinguished Paper Award at AdvML@CV CVPR 2026!

🌐 Project: https://t.co/JqGDhsNtvo

w/ @Samar_M_Fares @ziu_klea @tolusophy @TakacMartin @FuaPv 👇

[Nikita Durasov @norange_ai]

6/6 📊 The Results: MirrorCheck consistently outperforms baseline methods , achieving detection accuracies as high as 99% across diverse victim models and multimodal attack scenarios!

[Nikita Durasov @norange_ai]

4/4
📄 Paper: https://t.co/qpqDDTrNwM
📝 arXiv: https://t.co/v2jfGTqWQa
💻 Code: https://t.co/zvfuTMZLZt
🔬 Colab: https://t.co/Ds2SIiFdC1
🎬 Video: https://t.co/1ZbALLqRVO

#DiffusionModels #3DGaussianSplatting #TMLR #ComputerVision

[Nikita Durasov @norange_ai]

1/4🎉 Our paper "𝐃𝟑𝐃𝐑: Diffusion Models are Secretly Zero-Shot 3DGS Harmonizers" is in TMLR!

Diffusion models implicitly learn scene lighting — we leverage this for zero-shot 3DGS object insertion & relighting. No env maps. No material estimation.

🌐 https://t.co/wRwUbQDtND

[Nikita Durasov @norange_ai]

3/4 To preserve fine object details (prints, textures, patterns), 𝐃𝟑𝐃𝐑 introduces a new diffusion personalization technique that goes beyond standard DreamBooth.

+2.0 dB PSNR improvement over existing methods. We also release a new benchmark dataset to drive future research.

[Nikita Durasov @norange_ai]

@DamienTeney @FuaPv Hey @DamienTeney, thanks! Yes, the energy-based intuition is actually something we've been thinking about — it played a big role in how we came up with the idea for the method. I think it also nicely coexists with the manifold attractor interpretation!

[Nikita Durasov @norange_ai]

9/
Summary:

IT³ is simple, effective, and general.
It pushes predictions toward self-consistency, without needing task-specific tricks.

Try it on your own model model and data below.

📄 https://t.co/KTyBOK65ft
💻 https://t.co/7zaMGESAjU
🧠 https://t.co/uQRvmu5hok

[Nikita Durasov @norange_ai]

🧵 1/
Excited to share our #ICML2025 paper:
“IT³: Idempotent Test-Time Training”
A simple, domain-agnostic method to adapt models during inference, using only the test instance — no auxiliary task, no tuning. Let’s dive in 👇
🔗 https://t.co/pbvtKty4ch
https://t.co/gp7oilw6sS

[Nikita Durasov @norange_ai]

8/
The best part?

You can add IT³ to your model in minutes.
It’s implemented in torch-ttt and works with any architecture that accepts two inputs.

Check out the Colab:

🔬 https://t.co/9TLjrczVGy