Xinyin Ma (@horseeeMa) / X

Xinyin Ma

63 posts

Xinyin Ma

@horseeeMa

Ph.D. Candidate at National University of Singapore @NUSingapore. Interested in efficient learning for Diffusion Models, LLM, ... || Google PhD Fellowship 2024

Singapore

Joined August 2022

Xinyin Ma
@horseeeMa
Nov 15, 2024
Super excited to share that I’ve received the Google PhD Fellowship for 2024 🥳🥳🥳🥳🥳🥳 Heartfelt thanks to my advisor, @XinchaoWang3, and to all my incredible collaborators and friends who’ve supported me along the way. Also thanks @GoogleAI for this amazing recognition!
19K
Xinyin Ma
@horseeeMa
Dec 19, 2023
Our new work: 0.1% Data Makes Segment Anything Slim Page: huggingface.co/papers/2312.05… Github: github.com/czg1225/SlimSAM Compared to SAM-H, SlimSAM achieves approaching performance while reducing parameter counts to merely 0.9%(5.7M), MACs to 0.8%(21G), and requiring only 0.1% data.
Paper page - 0.1% Data Makes Segment Anything Slim
From huggingface.co
14K
Xinyin Ma
@horseeeMa
Jun 20, 2024
Hi everyone! I will present our work DeepCache today at @CVPR in the afternoon poster session! POSTER: Arch 4A-E Poster #111 Come stop by our poster today😀! We are very welcome for discussions. You can also drop me an email if you want to discuss more.
2.5K
Xinyin Ma
@horseeeMa
Jun 5, 2024
🥳Excited to share our new work: Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching. Our method nearly losslessly removes ~46% of the layers without updating model parameters. - paper: arxiv.org/abs/2406.01733 - code: github.com/horseee/learni…
2.6K
Xinyin Ma
@horseeeMa
Dec 20, 2023
Thanks @_akhaliq for sharing our work! DeepCache is a training-free and almost lossless acceleration of diffusion models. Check our Github for more results😉:
AK
@_akhaliq
Dec 5, 2023
DeepCache: Accelerating Diffusion Models for Free paper page: huggingface.co/papers/2312.00… Diffusion models have recently gained unprecedented attention in the field of image synthesis due to their remarkable generative capabilities. Notwithstanding their prowess, these models often
GitHub - horseee/DeepCache: [CVPR 2024] DeepCache: Accelerating Diffusion Models for Free
From github.com
12K
Xinyin Ma
@horseeeMa
Dec 20, 2023
DeepCache is integrated into OneDiff, and DeepCache with OneDiff can achieve 3x speed up😎 See Github: github.com/Oneflow-Inc/on… and Reddit: reddit.com/r/StableDiffus… Big Thanks for Oneflow and OneDiff!
4.3K
Xinyin Ma
@horseeeMa
Feb 27, 2024
Excited to announce that DeepCache got accepted to #CVPR2024! 🎉 Huge thanks to @XinchaoWang3 and Gongfan! 🍊Paper: arxiv.org/abs/2312.00858 🦒Code: github.com/horseee/DeepCa… 🐳 Diffusers: huggingface.co/docs/diffusers…
AK
@_akhaliq
Dec 5, 2023
DeepCache: Accelerating Diffusion Models for Free paper page: huggingface.co/papers/2312.00… Diffusion models have recently gained unprecedented attention in the field of image synthesis due to their remarkable generative capabilities. Notwithstanding their prowess, these models often
1.4K
Xinyin Ma
@horseeeMa
Nov 28, 2024
Introducing CoDe: our new work on accelerating VAR! Exciting observations: 1. Downsizing models for larger scale has minimal impact on performance! 2. Different scales serve unique roles, so a unified model for all scales is not optimal. Check out our paper and code🥰
CZG1225
@chen_zigen56940
Nov 28, 2024
Excited to introduce our new work! CoDe: Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient CoDe achieves 1.7x speedup and 0.5x Memory with only a negligible impact to quality! 🚀 GitHub: github.com/czg1225/CoDe 🚀 Project Page: czg1225.github.io/CoDe_page/
670
Xinyin Ma
@horseeeMa
Jan 5, 2024
A doc page for DeepCache has been added in Diffusers! Huge thanks to the Diffusers team🥰 Check here for more information:
DeepCache · Hugging Face
From huggingface.co
717
Xinyin Ma
@horseeeMa
Dec 19, 2023
Replying to @horseeeMa
(1/3) We introduce SlimSAM, a novel SAM compression method that achieves superior performance with remarkably low training costs. This is achieved by the efficient reuse of pre-trained SAMs through a unified pruning-distillation framework.
252
Xinyin Ma
@horseeeMa
Jun 25, 2024
SO COOL!!!!!
Xingyi Yang
@yxy2168
Jun 25, 2024
🚀 Introducing Video-Infinity! Our new distributed framework revolutionizes long video generation. 🎥✨ 🌟 Generate videos up to 2,300 frames in just 5 minutes—100x faster than previous methods!#AI #Video #AIGC Project Page: video-infinity.tanzhenxiong.com Paper: arxiv.org/abs/2406.16260
00:00
00:00
00:00
00:00
426
Xinyin Ma
@horseeeMa
Dec 19, 2023
Replying to @horseeeMa
Arxiv:
arxiv.org
SlimSAM: 0.1% Data Makes Segment Anything Slim
Current approaches for compressing the Segment Anything Model (SAM) yield commendable results, yet necessitate extensive data to train a new network from scratch. Employing conventional pruning...
293
Xinyin Ma
@horseeeMa
Dec 20, 2023
Introducing another new work: DeepCache Paper: arxiv.org/abs/2312.00858 Github: github.com/horseee/DeepCa…… 🚀 Training-free and almost lossless 🚀Accelerate Stable Diffusion by 2.3X and Stable Diffusion XL by 2.6X😀 🚀Compatible with sampling algorithms like DDIM and PLMS
GIF
364
Xinyin Ma
@horseeeMa
Dec 19, 2023
Replying to @horseeeMa
(3/3) SlimSAM yields significant performance improvements while demanding over 10X less training costs than existing methods. Compared to SAM-H, SlimSAM achieves approaching performance while reducing parameters to merely 0.9%, MACs to 0.8%, and requiring only 0.1% (10k) data.
255