Chenlin Meng
@chenlin_meng
Followers
8K
Following
2K
Media
15
Statuses
252
Co-founder & CTO @pika_labs | ex @StanfordAILab @Stanford
Joined March 2020
Pika 2.1 is available to public! It supports 1080p resolution, sharp details, smooth/dynamic motion, lifelike human characters and many more. Want to unlock more potential, test it out yourself at ๐๐.
Getting really real with youโฆ. Pika 2.1 is HEREโcrystal-clear 1080p resolution, razor-sharp details, seamless motion, lifelike human characters, and more. If you see it, youโll believe it. Try it out now at pika dot art.
5
0
38
Super excited for the launch of Pika 1.0! I am extremely grateful to be working with such an amazing and talented team on this journey! โค๏ธ. I am also very thankful for the support from our incredible investors, advisors, friends, and community members. We couldn't have achieved.
Introducing Pika 1.0, the idea-to-video platform that brings your creativity to life. Create and edit your videos with AI. Rolling out to new users on web and discord, starting today. Sign up at
53
35
620
Excited to share our work "On distillation of guided diffusion models"! Our distillation approach allows classifier-free guided diffusion models to generate high-quality samples using as few as 1-4 sampling steps๐ฎ
10
87
573
Excited to share our work SDEdit @YSongStanford, @baaadas, @jiajunwu_cs, @junyanz89, @StefanoErmon. Based on stochastic differential equations, SDEdit allows various image editing and synthesis applications without task specific training and optimization.
2
56
254
Welcome @omerbartal ๐๐ Looking forward to the amazing journey ahead! ๐ฆพ.
Thrilled to share that I've joined @pika_labs as Founding Scientist!. It's exciting times in video generation and AI, and we're committed to unlock new capabilities and empower creative content creation. Delighted to work with the most talented @demi_guo_ @chenlin_meng๐.
33
112
113
Score matching has proven effective in modeling continuous data modalities. However, this approach is not applicable in discrete domains where the gradient is undefined. We propose an analogous score function called the โConcrete scoreโ for discrete data. #NeurIPS2022
2
15
153
Super excited to announce our new round led by @sparkcapital !! Extremely grateful for our amazing investors and the entire Pika team!! Really lucky to work with all of you โค๏ธ . Can't wait for the new journey ahead ๐ฆพ.
Excited to announce our $80M Series B funding led by @sparkcapital. Grateful to have our investors and team members joining us along this journey - can't do this without you all. Stay tuned for more updates from us later this year! ๐. Read more:
18
5
141
Excited to share our #ICLR2023 paper "Dual Diffusion Implicit Bridges for Image-to-Image Translation" (DDIB), an image-to-image translation approach with diffusion models ๐ฑ -> ๐ฏ .
1
26
131
We will be presenting our work SDEdit at #ICLR2022 .Tue 26 Apr 6:30 p.m. PDT - 8:30 p.m. PDT! Come to our poster session today and see how SDE-based generative models can be used for guided image synthesis and editing! ๐จ๐ฉโ๐จ.
1
4
52
Interested in making diffusion models fast? Iโll present our paper On Distillation of Guided Diffusion Models and other interesting works at #NeurIPS workshop on score-based methods this afternoon. ๐/๐: room 293-294 / 2:30pm - 2:40pm and 3:00pm-4:30pm.
Excited to share our work "On distillation of guided diffusion models"! Our distillation approach allows classifier-free guided diffusion models to generate high-quality samples using as few as 1-4 sampling steps๐ฎ
2
3
50
Super excited to share our update on PIKA 1.5. Check out our mind-blowing Pikaffects ๐คฏ.
Sry, we forgot our password. PIKA 1.5 IS HERE. With more realistic movement, big screen shots, and mind-blowing Pikaffects that break the laws of physics, thereโs more to love about Pika than ever before. Try it.
2
0
54
Thank you for covering our story @kenrickcai @Forbes ๐.
Thanks @kenrickcai from @Forbes for covering our story!.
7
2
43
Thank you so much for everyone's contribution to our tutorial! ๐๐๐ There are so many amazing works but we are only able to cover part of them ๐ขHere are some more papers Please feel free to add any related works to the repo!! โค๏ธ.
Sharing a list of very useful references from @chenlin_mengโs wonderful tutorial on the latest diffusion models for text-to-image synthesis, semantically-guided image editing, and controllable generation! ๐๐๐
๐๏ธ๐๐. #CVPR2023
0
6
41
@YSongStanford @baaadas @jiajunwu_cs @junyanz89 @StefanoErmon ๐จ More editing results with SDEdit, an image synthesis and editing framework based on stochastic differential equations @StefanoErmon ๐๐ฒ. Joint work with @YSongStanford @baaadas @jiajunwu_cs @junyanz89 @StefanoErmon . ๐
0
8
40
This approach is also effective for video generation and has been used in #ImagenVideo: it can distill video diffusion models down to just 8 sampling steps without noticeable loss in perceptual quality! . Work with @RuiqiGao, @dpkingma, @StefanoErmon, @hojonathanho, @TimSalimans.
1
3
40
Here are some ImageNet samples using only one sampling step! .(w denotes the guidance strength for a classifier-free guided diffusion model)
1
2
38
Super grateful for our incredible community, our amazing investors, and the entire Pika team!!โค๏ธ.
Itโs been a hell of a year. And today weโve capped it off with a humbling fundraise, led by @sparkcapital. Huge thanks to our incredible community. We wouldnโt have gotten here without you!
4
4
36
Our #ICML2023 Workshop on Structured Probabilistic Inference & Generative Modeling will take place at Hawaii Convention Center, Room 323 on Friday, July 28. Come to our workshop to learn more about generative modeling and meet our amazing speakers and panelists! ๐โบ๏ธ
2
4
31
Welcome @EladRichardson to the Pika Family!! ๐ฅณโบ๏ธ So excited! @pika_labs.
Very excited to share that I'm joining @pika_labs as a research scientist ๐. Great things ahead! ๐ฝ๏ธ
11
0
15
We also explore distilling diffusion models with guidance for style transfer applications. Orange๐ -> Bell pepper ๐ซ (16 sampling steps).Orange ๐ -> Acorn squash ๐ (16 sampling steps).(w denotes the guidance strengths)
1
0
26
Excited to share our NeurIPS 2020 paper on training autoregressive score models:. joint work with @yulantao1996, @YSongStanford, @baaadas and @StefanoErmon.
0
5
26
Come to our poster session on Thu, Dec 1 at 09:00 AM -- 11:00 AM (PST) @ Hall J 115 #NeurIPS2022.Paper: Joint work with @kristyechoi @baaadas @StefanoErmon.
0
3
23
Our approach works well for a wide range of guidance strengths using only one distilled model per dataset. We also develop a stochastic sampler for the distilled model. We observe that the stochastic sampler often outperforms the deterministic sampler in terms of Inception scores
1
1
22
@pika_labs will be at ICLR! Ping us if youโre there #ICLR2024 ๐ฅณ Our swags will never let you down โบ๏ธ.
Flying tomorrow to ICLR! .Happy to chat about video generation, diffusion models, and @pika_labs !. (this bag contains only Pika swag)
2
1
21
Organizing our workshop on Structured Probabilistic Inference and Generative Modeling at ICML 2023 ๐.If you are interested in reviewing and learning the frontiers of probabilistic inference and generative models ๐ช, then please fill out the form below๐.
Interested in reviewing and learning the frontiers of probabilistic inference and generative models? Sign up this form for ICML 2023 workshop Structured Probabilistic Inference and Generative Modeling! #ICML2023.
0
2
18
Here is the schedule of our #SIGGRAPH2023 Course on Diffusion Models with the amazing .@baaadas,@ShuangL13799063,@junyanz89, @karsten_kreis, @chenhsuan_lin, @TsungYiLinCV, @StefanoErmon ๐.Looking forward to seeing everyone tomorrow in Room 403 AB! ๐.
@siggraph @chenlin_meng @baaadas @ShuangL13799063 @junyanz89 @chenhsuan_lin @TsungYiLinCV @StefanoErmon Schedule:.๐ธ9:00am-10:00am Introduction to Diffusion Models.๐ธ10:00am-10:30am Latent Diffusion Models and Their Applications.๐ธ10:30am-11:00am Compositional Visual Generation.๐ธ11:00am-11:30am Customizing Diffusion Models.๐ธ11:30am-12:00am Diffusion Models for 3D Asset Generation.
1
1
13
Come to our ICML workshop on July 28! โบ๏ธ๐๏ธ.
Don't miss out on our SPIGM workshop ( at #ICML2023 in Hawaii๐๏ธ next week! . For better communication, please feel free to join our Slack channel: to connect with like-minded individuals and say hi to the community๐ฉโ๐ป!.
0
2
10
@AravSrinivas Thank you so much Aravid! Super grateful to have you as our early investor. Really thankful for your support!.
0
0
7
Work led by Xuan Su, an amazing master's student at Stanford ๐. Joint work with @baaadas and @StefanoErmon .[2/7].
1
0
6
Welcome @yuvalalaluf !! Looking forward to the amazing journey ahead ๐.
Very happy to share that I've joined @pika_labs as a Research Scientist! ๐. I'm excited about this new adventure and the opportunity to work with such an amazing team! ๐ฝ๏ธ๐
2
0
6
Empirically, we demonstrate the efficacy of CSM on density estimation tasks on a mixture of synthetic, tabular, and high-dimensional image datasets, and demonstrate that it performs favorably relative to existing baselines for modeling discrete data.
1
0
6
My amazing collaborators @robrombach @RuiqiGao will be presenting our work at @CVPR ! @robrombach will also be giving the Award Candidate Talk this Thursday afternoon. Feel free to stop by and ask any questions! Looking forward to the discussion! ๐.#CVPR2023.
1
0
5
โConcrete scoreโ is a generalization of the (Stein) score for discrete settings. Given a predefined neighborhood structure, the Concrete score of any input is defined by the rate of change of the probabilities with respect to local directional changes of the input.
1
0
5
@natfriedman Super grateful for your support from Day 1! We are really lucky to have you as our investor โบ๏ธ Couldn't ask for more.
0
0
4
This formulation allows us to recover the (Stein) score in continuous domains when measuring such changes by the Euclidean distance, while using the Manhattan distance leads to our novel score function in discrete domains.
1
0
4
Our proposed approach DDIB solves both problems:.โ
It does not require diffusion models to be trained jointly on two domains. โ
It can be easily adapted to other domain pairs by leveraging pre-trained off-the-shelf diffusion models. [4/7].
1
2
4
Our distillation approach allows classifier-free guided diffusion models to generate high-quality samples using as few as 1-4 sampling steps ๐ It is also effective for data modalities beyond images ๐ฎ๐.
1
0
4
1
0
3
Current image translation methods such as GANs:.โ๏ธ require the presence of both datasets during training, which betrays data separation and data privacy. โ๏ธ๏ธcannot be easily adapted to translations between other source-target pairs. [3/7].
1
0
3
Joint work with my amazing collaborators and advisors .@electronickale @YSongStanford @baaadas @jiajunwu_cs @junyanz89 @StefanoErmon๐.
0
0
3
0
0
3
Based on stochastic differential equations, SDEdit allows various image editing and synthesis applications without task-specific training and optimization.
1
0
3
Finally, we introduce a new framework to learn such scores from samples called Concrete Score Matching (CSM), and propose an efficient training objective to scale our approach to high dimensions.
1
0
3
Theoretically, we can interpret the translation procedure as two concatenated Schrรถdinger Bridges: an optimal transport process where the diffusion models minimize the pixel-wise distance between the translated images (see animation below).[6/7]
1
0
3
We provide a wide variety of results on both synthetic and ImageNet datasets. Check out the details in our paper and project page!.[7/7].
1
0
2
DDIBs rely on so-called probability flow (PF) ordinary differential equations (ODEs). The forward ODE obtains latent code from the source image, while the reverse ODE then constructs the target image. [5/7]
1
0
2
Joint work with @robrombach, @RuiqiGao, @dpkingma, @StefanoErmon, @hojonathanho, @TimSalimans .@StanfordAILab @StabilityAI @GoogleAI.
0
0
2
0
0
2
0
0
1
@tri_dao @pika_labs Thank you for your support! โบ๏ธ Really fortunate to collaborate with you tooโค๏ธ.
1
0
1
Come to our SIGGRAPH course on Diffusion Models tomorrow!! ๐๐.
Come to our SIGGRAPH course on Diffusion Models tomorrow (Thu 9 am-noon, Room 403AB).
1
0
1