Yigit Ekin

Yigit Ekin

ygtekn01@gmail.com

I am a Research Scientist on the post-training team at fal.ai, where I work on large-scale diffusion models for image and video generation. Previously, I was a research intern at Reve AI working on multimodal vision language models. I received my M.S. in Computer Engineering from Bilkent University, where I was supervised by Aysegul Dundar. My research focuses on generative computer vision, with an emphasis on diffusion models for controllable video/image generation and editing. Previously, I worked on vision-language models and interpreting the inner workings of deep neural networks.

Publications

  1. BeyondMasks benchmark examples showing causal and physical effects in video object removal
    BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal
    ECCV 2026
    Yigit Ekin, Enes Sanli, Aykut Erdem, Erkut Erdem, Aysegul Dundar
    TL;DR. A benchmark for testing whether video object removal methods also remove causal and physical after-effects, including shadows, reflections, illumination changes, translucency, steam, and dynamic traces.
    PDF Project Page Dataset Code
  2. Diffusion Sliders paper teaser
    The Unreasonable Effectiveness of Text Embedding Interpolation for Continuous Image Steering
    Preprint
    Yigit Ekin, Yossi Gandelsman
    TL;DR. A training-free method that steers text-conditioned image and video generators continuously by adding automatically constructed semantic directions directly to their text embeddings.
    PDF Code Project Page
  3. RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers
    NeurIPS 2025
    Ahmet Berke Gokmen*, Yigit Ekin*, Bahri Batuhan Bilecen*, Aysegul Dundar
    * denotes equal contribution
    TL;DR. A training-free video motion transfer method that encodes reference motion into diffusion transformers by warping and optimizing their rotary positional embeddings.
    PDF Code Project Page
  4. CLIPAway paper teaser
    CLIPAway: Harmonizing Focused Embeddings for Removing Objects via Diffusion Models
    NeurIPS 2024
    Yigit Ekin, Ahmet Burak Yıldırım, Erdem Eren Çağlar, Aykut Erdem, Erkut Erdem, Aysegul Dundar
    TL;DR. A plug-and-play object removal method that uses background-focused CLIP embeddings to prevent diffusion inpainting models from hallucinating the removed object.
    PDF Code Project Page
  5. MixGAN paper teaser
    MixGAN: Dual Path StyleGAN Fusion for Diverse and Editable Inpainting
    Preprint
    Mustafa Utku Aydogdu, Ahmet Burak Yıldırım, Yigit Ekin, Aysegul Dundar
    TL;DR. A dual-path StyleGAN framework that mixes encoded-image and random-sample features to produce diverse, high-fidelity inpaintings while preserving editability.

Research Experience

  • Work on large-scale diffusion models for image and video generation.
  • Designed exploratory experiments for internal LLM and VLM variants across training configurations.
  • Implemented a speculative-decoding training pipeline for internal models.
  • Curated million-scale multimodal and text datasets for pretraining and supervised fine-tuning.
  • Investigated data mixtures and multimodal fine-tuning strategies while preserving language performance.

Education

Supervised by Aysegul Dundar · Departmental Scholarship

Teaching

Teaching Assistant

Bilkent University

  • CS 224: Computer Organization - Spring 2024 and 2025
  • CS 484: Introduction to Computer Vision - Fall 2024
  • CS 101: Algorithms and Programming I - Fall 2023

Best Teaching Assistant Award, 2025