Tahira Kazimi

Intern @Apple | PhD student @virginia_tech | ex @METU_ODTU | Avid reader

[Tahira Kazimi @TKazimi415]

Diverse Video Generation with DPP-Guided Policy Optimization is accepted to #CVPR2026 ✨

We tackle diversity in T2V models by formulating diversity as a set-level optimization problem.

Huge thanks to my advisor @PINguAR and my collaborator Connor Dunlop

[Tahira Kazimi @TKazimi415]

✨We introduce Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization: a framework for diverse video generation that combines Determinantal Point Processes and GRPO theories to enforce explicit reward on diverse generations.

[The Sanghani Center at Virginia Tech @SanghaniCtrVT]

#CVPR2026 @TKazimi415 presented collaborative work w/Connor Dunlop, fellow Ph.D. student @SanghaniCtrVT, & advisor Pinar Yanardag on a framework that enables text-to-video models to produce diverse yet prompt-faithful videos from a single input prompt.
https://t.co/cWxWP5n30t

[Pinar Yanardag @PINguAR]

Visit my amazing PhD student Tahira @TKazimi415 presenting her poster on diverse video generation at @CVPR #CVPR

[Pinar Yanardag @PINguAR]

Join us on this morning from 8:30AM–12:30PM for the 2nd Personalization in Generative AI Workshop (P13N) at @CVPR! #CVPR2026 #P13N We have an amazing speaker and panelist lineup including @natanielruizg @AbermanKfir @OPatashnik @RanaHanocka @ElorHadar More details: https://t.co/EdVQpqYOxa

[Hidir Yesiltepe @d_yesiltepe]

✨ We introduce VideoMLA, an MLA-style latent KV cache for autoregressive video diffusion.

Multi-Head Latent Attention (MLA) has been central to the efficiency of DeepSeek-V2 and V3. In this work, we study whether the same principle can be transferred from language modeling to long-horizon video generation.

[Kiymet Akdemir @akdemir_kiymet]

Game engines let you spawn anything into a scene. World models don't. Once the camera moves, you're stuck with whatever the model dreamed up.

🧵 We fix that with SPAWN, a training-free method that injects a custom concept into the rollout from either an image or a text prompt.

[Tahira Kazimi @TKazimi415]

Check out our poster at #CVPR2026
Friday, Jun 5, 4:00 - 6:00 PM MT
ExHall A & F #526

[The Sanghani Center at Virginia Tech @SanghaniCtrVT]

Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization
Main proceedings
@TKazimi415 Connor Dunlop @PINguAR (all VT)
https://t.co/cWxWP5n30t
TL;DR: Enables text-to-video models to produce diverse yet prompt-faithful videos from a single input prompt.

[Pinar Yanardag @PINguAR]

GEMLAB is heading to #CVPR2026! We’re excited to share our papers and talks across controllable generative AI, long-horizon video modeling, and personalization! #CVPR2026 #CVPR

Full schedule: https://t.co/To0i22tYog

[Pinar Yanardag @PINguAR]

Proud to share: GEMLAB has 3 main-track and 1 Findings accepted to #CVPR2026! More details soon. Huge congrats to my amazing students & collaborators. See you in Denver!

P.S. GEMLAB is keeping the tradition alive 3 years in a row with 3 papers at every CVPR 🚀 https://t.co/L8GfK7kAeE

[Pinar Yanardag @PINguAR]

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 💎

[Tahira Kazimi @TKazimi415]

Thanks to my collaborator Connor Dunlop, and our advisor @PINguAR for her invaluable insights.
Check out our project page!
🌐 Project Page: https://t.co/cwjQU8soGw
Paper: https://t.co/4easn51Bg2
Huggingface: https://t.co/T4R5BUrskU

[Tahira Kazimi @TKazimi415]

✨We introduce Diverse Video Generation with Determinantal Point Process-Guided Policy Optimization: a framework for diverse video generation that combines Determinantal Point Processes and GRPO theories to enforce explicit reward on diverse generations.

[Tahira Kazimi @TKazimi415]

Our method improves video diversity on Wan2.1 and CogVideo, enhancing motion, temporal variation, and scene diversit, while keeping CLIP fidelity stable. It also raises video-quality scores, showing DPP-GRPO boosts diversity without sacrificing alignment.

[Tuna Meral @tunahansalih]

This dog + its favorite toy = one perfect image.

That’s the test. Can your #GenAI model combine two unique subjects without errors? Show us.

[Tahira Kazimi @TKazimi415]

Check out our work and let us know what you think!
Project Page: https://t.co/dSkaLHCxyN
Arxiv: https://t.co/XTxTudNx54
HuggingFace: https://t.co/lGyjgg9i9u

[Tahira Kazimi @TKazimi415]

We introduce Audit & Repair, a collaborative multi-agent framework that autonomously identifies, corrects, and refines inconsistencies across multi-panel story visualizations.
This is a joint work with @akdemir_kiymet, supervised by @PINguAR

[Tahira Kazimi @TKazimi415]

Our Audit & Repair framework achieves best-in-class consistency in story visualization:
CLIP-I ↑ 0.850, DINO ↑ 0.568, LPIPS ↓ 0.472, HPS ↑ 0.319, TIFA ↑ 0.71, outperforming StoryDiffusion, StoryGen, ConsiStory & others across all major metrics.