|
Hayk Poghosyan I'm a Senior ML research engineer at Pipio, where I work on generative AI for audio-conditioned video generation. Check out our latest project: EditYourself. Previously, I was a senior ML scientist at PicsArt, focusing on large-scale controllable image and video generation using GANs and diffusion models, low-light enhancement and person segmentation. I've also worked on AI website builders, time-series forecasting, and anomaly detection. I have 7+ years in experiance in both research and engineering. I hold a PhD in theoretical physics from the Yerevan Physics Institute, with research spanning classical and quantum spin systems, and a master's degree in informatics and computer engineering. My work bridges computer vision, generative models, and large-scale data systems, with publications at venues including CVPR, JHEP, ICASSP, ACM Multimedia, and Physica A. Email / LinkedIn / Scholar / GitHub đź“Ť Alicante, Spain |
|
Experience |
|
ResearchI'm interested in computer vision, deep learning, generative AI, and image processing. Most of my work focuses on image and video generation, controllable image editing, and inferring structure and semantics from visual data using modern generative models as well as optimizing them for real world loads and usecases. |
|
EditYourself: Audio-Driven Generation and Manipulation of Talking Head
Videos with Diffusion Transformers
NEW
John Flynn, Wolfgang Paier, Dimitar Dinev, Sam Nhut Nguyen, Hayk Poghosyan, Manuel Toribio, Sandipan Banerjee, Guy Gafnistrong> ArXiv 2026 project page · arXiv EditYourself is a diffusion-based video editing model for talking heads, enabling transcript-driven lip-syncing, insertion, removal and retiming of speech while preserving identity and visual fidelity. |
|
StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation
from Text
Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, Humphrey Shi CVPR 2025 project page · arXiv 
An autoregressive approach for long video generation (80 to 1200+ frames) with temporal consistency, high motion dynamics, and smooth transitions. Uses a conditional attention module (CAM) and appearance preservation module (APM); includes StreamingSVD (image-to-video) and StreamingModelscope (up to 2 minutes). |
|
Grounded-Instruct-Pix2Pix: Improving Instruction Based Image Editing with
Automatic Target Grounding
Artur Shagidanov, Hayk Poghosyan, Xinyu Gong, Zhangyang Wang, Shant Navasardyan, Humphrey Shi code 
A robust framework for localized instruction-based image editing. Two stages: (1) grounding mask extraction via CLIP-Score Filtering and Grounded-SAM, (2) localized editing with Instruct-Pix2Pix and latent blending. Incurs little overhead and requires no additional user inputs. |
|
ReCoRo: Region-Controllable Robust Light Enhancement by User-Specified
Imprecise Masks
Dejia Xu, Hayk Poghosyan, Shant Navasardyan, Yifan Jiang, Humphrey Shi, Zhangyang Wang ACM Multimedia 2022 code · paper (PDF) 
A low-light enhancement method that lets users specify where and how much to enhance an input image. Robust to roughly-supplied user masks; supports both imprecise and fine matting masks. |
Publications |
|