Hugo Caselles-Dupré

Honored to have been selected among the Forbes 30 under 30 with my fellows at @obv_ious 🎉🚀🙏 #2068e12e6eb1" target="_blank" rel="nofollow noopener">forbes.com/profile/obviou…

Merci Céline Calvez pour cette invitation à l'Assemblée Nationale, c'est un honneur de pouvoir partager notre expérience artistique et de recherche dans le cadre de la mission IA et Connaissances ! Lien en commentaire.

Seeing these ideas converging across labs (academia and industry) confirms something we felt early: interactive generative video will reshape art, games, simulation, and how we think about creation in time.

Portal runs around 50 FPS and has been deployed live at Parc des Princes, at the AI Summit in the Grand Palais, and other public venues. It’s not a clip you play, it is a video that lives and reacts.

A strong pattern is emerging across recent video generation papers. Different methods, same direction: video models are becoming real-time, autoregressive, and interactive. A quick thread on how this space is blooming. 📸: Rolling Forcing method.

Une video de notre intervention pendant #WAFT2025 organisé par @Bpifrance, pour parler de notre projet Molière Ex Machina! youtube.com/watch?v=HGgFnK…

Merci The New Siècle pour l'interview, à retrouver sur leur site : thenewsiecle.fr/interview-hugo…

Je serai en direct sur RFI aujourd'hui à 18h30 pour le débat du jour sur la création artistique avec l'IA. Suivez le lien pour écouter en direct ! rfi.fr/fr/direct-monde

Jean Zay est un ordinateur iconique pour les chercheurs en IA, que j'utilise personnellement depuis des années. Avec Obvious, on est fiers de présenter cette oeuvre-hommage qui sert de covering à la quatrième extension de la machine.

Liens : nouvelobs.com/economie/20251… franceinfo.fr/culture/specta…

Notre projet Molière Ex Machina est dans le Nouvel Obs et France Info ! Ravi de l'envergure qu'est en train de prendre cette aventure artistique et scientifique initiée par SU et Obvious.

@peterom Thanks for sharing. Definitely agree with you! It's been on my todo list as well but indeed, training a high quality IPA is a lot of work. Let's hope IPAs make a return this year🤞

Yeah, I trained one for Kontext also and it wasn't worth sharing Imo IP-Adapters are a fundamentally better approach but I lack the expertise to train one so have been making do hacking on top of other models. I do think there's a LOT we can do with edit models but it feels like an inferior approach. I think an IP-Adapter on a modern model like Qwen or Wan 2.2 could do a bunch of tasks with fine-tuning very easily - characters, styles, scenes, etc. - and the work of @cubiq & @ostrisai on SDXL/SD1.5 showed us how elegantly different functions could potentially combine I also believe that they could hugely enhance the animation flexibility to Wan 2.2 - to make it interpret animated images more naturally as Midjourney does - so I would probably bias towards a Wan 2.2 one. I'm pretty sure that's the secret sauce behind why MJ Video animates so well. We're also lucky enough to ahve a 512x512px SIGLIP lately that could be applied to them

For the whole time Flux was dominant in the open source ecosystem, its inability to do basically any style has been so frustrating for me Midjourney does great style references but I’m too much of a fanatic to enjoy it Qwen is such a breath of fresh air ❤️‍🔥

Large-scale composite images experiments

Excited for our new exhibition!

Fresco study N+1

Fresco study N