CS PhD @RiceCompSci @vislang | Research Intern @Snap | GenAI & Video & Audio & Computer Vision
Joined January 2025
- Tweets58
- Following145
- Followers115
- Likes186
Moayed Haji Ali retweeted
Introducing ActionParty: the first video world model that controls up to 7 players simultaneously on the same screen across 46 game environments.
We tackle the action binding problem in video diffusion, ensuring each player's action is applied to the right subject. 🧵
Moayed Haji Ali retweeted
🎉EgoEdit @Snapchat has been accepted to CVPR 2026! 🏆👻
We are bringing high-quality, real-time editing to egocentric videos. Our massive 100k video dataset and benchmark are ALREADY PUBLIC! 🔓🚀
🏠 Project Page: snap-research.github.io/EgoE…
🤗 Dataset: huggingface.co/datasets/ligu…
Moayed Haji Ali retweeted
Same problem in motion diffusion: a standing pose needs almost no denoising while a complex hand-object interaction needs everything the model has. Nobody's done elastic compute for motion DiTs yet. Easy win waiting to happen.
xitter.cf/moayedhajiali/status/2…
Not all pixels are equally hard, but DiTs still allocate compute uniformly across pixels, wasting efforts on easy regions. ELIT adds two lightweight cross-attention layers to focus compute where it matters, cutting FID by 53%.
ELIT: snap-research.github.io/elit
Not all pixels are equally hard, but DiTs still allocate compute uniformly across pixels, wasting efforts on easy regions. ELIT adds two lightweight cross-attention layers to focus compute where it matters, cutting FID by 53%.
ELIT: snap-research.github.io/elit
ELIT also comes with a nice perk: variable budgets give you a built-in weaker model. That gives autoguidance for free, and combined with class dropping it enables Cheap CFG, improving quality while cutting guidance cost by ~33% on Qwen-Image.
Huge thanks to all my collaborators @WilliMenapace @isskoro, @dogyunPark, Anil Kag, Michael Vasilkovsky, @SergeyTulyakov, Vicente Ordonez, @siarohin9013
ELIT will be presented in CVPR 2026
Project page: snap-research.github.io/elit
Code: github.com/snap-research/eli…
🚀 Excited to present Decomposable Flow Matching (DFM) tomorrow @ NeurIPS!
DFM introduces a lightweight DiT modification to unify progressive diffusion benefits, achieving up to 50% faster convergence.
📍 Poster: Friday, 4:30–7:30 PM — #4908
Webpage: snap-research.github.io/dfm/
Big news 📢📢
AV-Link has been accepted to ICCV 2025 🚀🚀
Can pretrained diffusion models connect for cross-modal generation?
📢 Introducing AV-Link ♾
Bridging unimodal diffusion models in one framework to enable:
📽️ ➡️ 🔊 Video-to-Audio
🔊 ➡️ 📽️ Audio-to-Video
🌐: snap-research.github.io/AVLi…
📄: hf.co/papers/2412.15191
⤵️ Results
Where are good old progressive diffusion models? 🤔
Breaking generation to multiple resolution scales is a great idea, but complexity (multiple models, custom diffusion process, etc) stalled scaling. Our Decomposable Flow Matching packs multi-scale perks into one scalable model.
Great for Image generation! 📷
When measuring performance with Fréchet DINO Distance (FDD) on ImageNet-1k 512px, DFM cuts the training compute by ~50% compared with the base architecture, and achieves 26.4% better than the best performing baseline.
Great for Image generation! 📷
When measuring the performance with Fréchet DINO Distance (FDD) on ImageNet-1k 512px, DFM cuts the training compute by ~50% compared with the base architecture, and achieves 26.4% better than the best performing baseline.
