🎬 2026 will be the year of autoregressive video models. As we wrap up 2025, we ask a critical question: How far can a diffusion model distilled with Self-Forcing on only 5-second, 16-FPS videos be pushed into long-form video generation without any supervision? ✨We introduce

So grateful for the incredible students and alumni of GEMLAB! Together, we published 4 main conference papers + 4 workshop papers at @NeurIPSConf. You’re all absolute gems 💎

Today we met the legendary @docmilanfar and in the back @yusuf_dalva looked terrified by his sudden appearance :)

We are live today with @d_yesiltepe ! Make sure to drop by to hear about LoRAShop, where we perform multi concept generation + editing without any training! 📍 Thu, Dec 4 · 4:30–7:30 PM PST 📷 Poster #4306 Project: https://t.co/YY7kHSfaXP Code:

My PhD student Hidir @d_yesiltepe is presenting his poster right now! @NeurIPSConf

📸 Presenting Dynamic View Synthesis as an Inverse Problem! 📍Exhibit Hall C, #5306 ⏱️11:00 AM - 2:00 PM

Thanks @_akhaliq for sharing our work. Self-Forcing unlocks ✨ real-time ✨ long-form action-controllable dynamic video generation when combined with Infinity-RoPE! 💫 We will be releasing the code next week, and I would be glad to chat about the project in person at NeurIPS!

✨I will be at #NeurIPS2025 in San Diego presenting 2 projects. 💬 Would be glad to chat on long-form video generation, video distillation, dynamic view synthesis and video super-resolution! • Dynamic View Synthesis as an Inverse Problem: https://t.co/8Ly2poDLoi • LoRAShop:

Introducing Canvas-to-Image (C2I): A new paradigm where you define all controls within a single RGB canvas. 🎨 We simplify complex generation into one intuitive interface. Place specific Identities, Poses, and Boxes to control exactly who appears, how they pose, and where they

@tunahansalih @akaan_akan @oktay_kaan @PINguAR @fal 🌐 Project Page: https://t.co/NCNSiZ7HIH 📝 Arxiv:

🎊I sincerely thank all my excellent team members @tunahansalih, @akaan_akan, @oktay_kaan for their hard work and my supervisor @PINguAR for her valuable feedbacks throughout the project. Special thanks goes to @fal for providing the resources that made this research possible.

🔜To introduce dynamic scene cuts and add cinematic variation, we use RoPE Cut. It applies a controlled discontinuous jump in the temporal RoPE coordinates, re-anchoring the current block to a new temporal location. This allows background, environment, or time-of-day to change

✂ To achieve fast prompt responsiveness, we introduce KV Flush. It renews the KV cache by keeping only two anchors, the global sink frame and the last generated latent frame, allowing new prompts to take effect immediately while preserving smooth local motion.

💫When the KV cache size is beyond the base model’s natural horizon, Block-Relativistic RoPE transitions into the semanticization regime. Instead of assigning unique temporal indices to very old frames, their positions are collapsed to a shared minimum index, turning them into

♾️At the core of our method is Block-Relativistic RoPE, which redefines temporal encoding as a moving frame of reference. New blocks are rotated relative to the base model’s horizon, earlier blocks are rotated backward, and relative geometry is preserved.

Excited to share that LoRAShop has been accepted at #NeurIPS25 as a Spotlight! 🎉 Huge thanks to my collaborator @d_yesiltepe and my advisor @PINguAR 🙏 Code will be released soon, looking forward to connecting with the community in San Diego!

And accepted at #NeurIPS2025! See you in San Diego.

Dinner at #CVPR turned into a GenAI think tank. 🍜 Veo3, personalization rants, and nonstop energy from amazing folks across Fal and Google. Couldn’t have asked for a better crew. @natanielruizg @gorkemyurt @d_yesiltepe @d_yesiltepe @yusuf_dalva

Dynamic View Synthesis as an Inverse Problem Contributions: • We identify and formalize the Zero-Terminal SNR Collapse Problem, showing that while zero terminal SNR schedules improve generation quality, they inherently break injectivity, preventing deterministic inversion and