Jacob Fang
Jacob Zhiyuan Fang

Jacob Zhiyuan Fang

Member of Technical Staff · Microsoft AI

I work broadly on visual generation—from images and videos to visual assets—using diffusion models, code, and agentic methods.

Google ScholarLinkedInX

MAI Image Series Models

I am a core contributor to the MAI Image Series Models, focusing on post-training. My work spans reward model design, vision-language model RL for prompt enhancement, RL for captioning models, and diffusion RL.

MAI-Image-2.6 generated vintage-style snapshot of a family in an apartment courtyard MAI-Image-2.6 generated scene of a woman and cat in warm light, with drawings on the wall MAI-Image-2.6 generated image of a weathered gas meter with Chinese signage
MAI-Image-2.6 generated café portrait of a person holding coffee MAI-Image-2.6 generated pop-up illustrations of miniature Tokyo, Shanghai, Bangkok, and Singapore
MAI-Image-2.6 · Generated examples · Click to enlarge · LinkedIn ↗

Model updates

  1. MAI-Image-2.6 · Foundry availability update: Microsoft confirmed general availability. Source ↗

  2. MAI-Image-2.6 · I2I ranked #1 for image editing on Artificial Analysis. Source ↗

  3. MAI-Image-2.6 · T2I ranked #2 on Text-to-Image Arena. Source ↗

  4. MAI-Image-2.5 launched at #3 for text-to-image and #2 for image editing on Arena.

  5. MAI-Image-2 launched, bringing MAI to #3 among text-to-image model families on Arena.

Original Arena leaderboard for MAI-Image-2.6. All model rows are preserved.
Overall ranking
Original Arena category comparison for MAI-Image-2.6.
By category

Arena release snapshot · Click an image to enlarge · Source ↗

Earlier Arena results

The 2.5 Preview debuted at #3 on Text-to-Image Arena, with a score of 1,254. This earlier release snapshot precedes the 2.6 chart above.

Original Arena leaderboard for MAI-Image-2.5 Preview. All model rows are preserved.
Overall ranking
Original Arena category comparison for MAI-Image-2.5 Preview.
By category

Earlier release snapshot · Click an image to enlarge · Source ↗

TikTok Video Generation

At TikTok, I contributed to video generation models and creative features, including AI Alive.

AI Alive public demo preview
AI Alive19 secShou ↗
ATI research demo showing motion guided by trajectory instructions
ATITrajectory control · 1:09Research project ↗
MAGREF research demo showing reference-guided video generation
MAGREFReference guidance · 1:29Research project ↗

Hover to preview · Click to expandSelect a demo to watch

Amazon Nova Models

At Amazon, I contributed to the Nova family of image and video generation models.

Amazon Nova product image showing a generated dinosaur in a teacup
Amazon Nova. Image and video generation models.Amazon’s December 2024 announcement ↗

Experience & education

  1. Mar 2026 — Present
    Microsoft AIMember of Technical Staff
  2. Jul 2024 — Mar 2026
    ByteDanceSenior Research Scientist
  3. Jun 2022 — Jul 2024
    AmazonApplied Scientist
  4. May 2021 — Jan 2022
    MicrosoftResearch Internship
  5. Aug 2020 — May 2021
    MicrosoftResearch Internship
  6. May 2020 — Aug 2020
    MicrosoftResearch Internship
  7. Sep 2016 — Jan 2017
    Chinese Academy of SciencesResearch Assistant

Ph.D., Computer Science · Arizona State UniversityComputer vision and multimodal representation learning

B.Eng., Computer Science · Southern University of Science and Technology（南方科技大学）

Reviewer: ICCV, CVPR, ECCV, NeurIPS, ICLR, ICML, ACL, EMNLP, SIGGRAPH, SIGGRAPH Asia, and TMLR.

Selected publications

Scroll to browse

Controllable video generation

Image generation and generalization

Representation and vision-language learning

Thoughts

Image preview
Video demo