Nicolas Sereyjol-Garros (@sg_nicolas) profile picture

PhD student at @valeoai and @ImagineEnpc

Paris, France
Joined January 2023
Photo by Nicolas Sereyjol-Garros (@sg_nicolas)Photo by Nicolas Sereyjol-Garros (@sg_nicolas)Photo by Nicolas Sereyjol-Garros (@sg_nicolas)Photo by Nicolas Sereyjol-Garros (@sg_nicolas)Photo by Nicolas Sereyjol-Garros (@sg_nicolas)Photo by Nicolas Sereyjol-Garros (@sg_nicolas)
1/n Our paper REPA-G is accepted at #NeurIPS2026! 🎉 I'll present it at the NeurIPS Paris site 🇫🇷. We introduce test-time conditioning of REPA-style diffusion models with visual features, with no extra training.
3
25
119
5,759
Nicolas Sereyjol-Garros retweeted
My first preprint, ScoreCLR, is out on arXiv! 🧵 Contrastive learning where a diffusion score network does the repulsion. Embeddings reach near-full rank (229/256 on CIFAR-10, vs 96 for SimCLR and 58 for LeJEPA) at linear cost in batch size. 📄 arxiv.org/abs/2611.18469
1
3
13
931
Nicolas Sereyjol-Garros retweeted
Vision Banana finetuned a closed model & shared no code, weights or data. Its report left many design choices open, so we worked them out ourselves. OVB: one open FLUX.2 Klein model for segmentation, depth & normals that still edits images. 🍌Super fun project w/ @FunAILab team
🍌 What if an image generator could see? @GoogleDeepMind's Vision Banana showed it can, but kept data, recipe and weights closed. So we rebuilt it fully open: ✨ FLUX.2 Klein 4B/9B finetune w/ public data only ✨ 73.9 vs Vision Banana’s 69.9 mIoU zero-shot on Cityscapes ✨We release code + 148 checkpoints for the community/ Aaand we innovate: 📈 We develop a new test-time scaling recipe we call “legend-ensembling”: this gives strong gains by applying more compute without retraining anything. 🔗 openvisionbanana.github.io 📄 arxiv.org/abs/2611.15844 A @FunAILab work from A. Jadon, @v_pariza, @FragileGoodwill, @jonaruthardt, @ASophiaKoepke, @abursuc and me.
5
5
32
1,532
Nicolas Sereyjol-Garros retweeted
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training. Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
18
57
10
526
41,661
Nicolas Sereyjol-Garros retweeted
🧩 Can you tell a real book from AI writing? We built a fun little game called fauxlio.app Two excerpts: one's a classic, one's AI, and your goal is to pick the real one. 5 new rounds a day, EN/FR, free, no sign-up. Tell @VictorTXing and me what you think! 🫶
2
2
2
6
599
Nicolas Sereyjol-Garros retweeted
What happens when you simply swap two videos? For Large Multimodal Models, the answer can change surprisingly often! Come join us at #EMNLP2026 this morning to learn more about our work. arxiv.org/pdf/2610.24786 @paolorotaphd @eliricci_
10
21
480
Nicolas Sereyjol-Garros retweeted
After MIRA on Rocket League, I was very curious to know what it takes to bring interactive world models to the real world. It starts with the video autoencoder: how, where and how much to compress. CADRE is our answer, and it's all open 👇
Many world models borrow their video autoencoder from text-to-video generators, and with it, a noticeable lag on every action. Today we release CADRE, open CAusal Diffusion REpresentation autoencoders that let real-time world models react fast, on a single GPU 🧵
1
5
22
640
Nicolas Sereyjol-Garros retweeted
Hey all, we have a new paper: Dino Forcing Flow Models: Do not denoise what you can predict. Tldr; Denoising image and predicting (not denoising) related deterministic information (like dino) provides a model with faster convergence! Read the paper here: arxiv.org/abs/2610.11751v1
3
30
2
160
13,472
3/n Condition on a full feature map, a masked region, a global average or a single patch. Sum potentials to compose concepts, e.g. an object from one image and a background from another. No extra parameters, no fine-tuning.
1
1
9
378
Nicolas Sereyjol-Garros retweeted
🏛️ LLM agents in a new environment repeat the same mistakes. Meet DAEDALUS: agent memory bootstrapped from tasks the agent writes itself, with no training tasks or human-written guidelines. It makes 7 agents from various model families succeed more often, in fewer turns 🧵⬇️
4
5
5
27
1,983
Nicolas Sereyjol-Garros retweeted
1/n 🚀 New paper 𝚆𝚑𝚎𝚛𝚎-𝙾𝙿𝙳 In on-policy self-distillation, the student learns from a privileged version of itself. In recent methods, privilege often comes from a better view of the image. What if, instead of a better view, the teacher knew where to look? 👀
1
13
2
39
3,098
Nicolas Sereyjol-Garros retweeted
some autoencoders like mimi in audio or va-vae distill representations in the latent space to make it easier to generate. these works assume that the teacher has a structure that is somewhat shared by the student, which is not necessarily the case!
Many generative models work in an autoencoder's latent space; They train faster and generate better when this space is well organized. One way to organize it is distilling a pretrained model into it. But how can it work when the two spaces have different structures?
1
3
7
356
Nicolas Sereyjol-Garros retweeted
(1/n) Can older lidar SSC models catch up with recent SOTA just by getting better inputs? At #ICIP2026 today and in our new journal extension, we add two cheap priors to existing networks, with no redesign. 📍 Poster #2752, Park Hall, 14:00–16:00 🇫🇮 🧵👇
1
6
1
16
4,049
Nicolas Sereyjol-Garros retweeted
Can a car learn to drive by playing against itself, using only its cameras? Self-play driving agents have relied on a shortcut: perfect knowledge of every vehicle's position and speed, even the ones they can't see. A real car never gets that. We take the shortcut away. 🧵
1
6
1
21
6,119
Nicolas Sereyjol-Garros retweeted
New state of the art in one-step image generation 🏆 One forward pass. The method is so simple it's almost unsettling: no teacher, no discriminator, no denoising trajectory 🧵
1
9
19
812
Nicolas Sereyjol-Garros retweeted
Language modeling is ambiguous. If you ask someone to caption an image, describe an audio clip, or to translate a sentence, you’ll get several valid outputs. In this work, we handle this with a winner-takes-all training objective that encourages diverse, plausible predictions.
Replying to @valeoai
1/ Multiple Choice Learning of Low-Rank Adapters for Language Modeling TL;DR: LoRA heads + Winner-Takes-All loss → diverse, plausible LLM outputs 📄 arxiv.org/abs/2507.10419 💻 github.com/Victorletzelter/L… @VLetzelter, H. Malard, M. Fontaine, G. Richard, S. Essid, @abursuc, @ptrkprz
1
3
1
9
1,442
Nicolas Sereyjol-Garros retweeted
⏰🚨 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗱𝗲𝗮𝗱𝗹𝗶𝗻𝗲 𝗲𝘅𝘁𝗲𝗻𝗱𝗲𝗱!! Still time to apply for the ELLIS Summer School on Autonomous Driving 🚗 📅 New deadline: Friday, June 12, 11:59 PM AoE More info and registrations 👉 pages.cvc.uab.es/ssadriving/
📣Super happy to announce the ELLIS | ELLIOT Summer School on Autonomous Driving 🚗 Apply now! 📅 September 15-18, 2026 📍Barcelona, Spain 🌐 pages.cvc.uab.es/ssadriving/ @ELLISforEurope @ELLISBarcelona @CVC_UAB @valeoai
2
6
1,132