Richard Zhang @rzhang88 Twitter profile

Last Seen Profiles

@insideeducation

@M_Poniedzialek

@UkaszKremer

@hanni196286

@TheChemist

@carimcim1

@tx_coffee

@Zaydan0077

@Lyney777

@xxx_chiaki_08

@stelmet_sa

@heizoutistic

@slevkupectvi

@PiotrDudek_

@maciekmolka

@lax_fresh_flect

@ban__nasr

@supaagluee

@ObitsNFT

@GayCruisingBKK

@FrankLaddSTL

@A_llwnlnw

@vvalentique

@yasintim139

@UScavetta94905

@Audeze

@MomonoaZoro

@DrProctorOff

@TROYCETV

@POWERCOURT_09

@Analternateacc2

@AvidReadersCafe

@Koz112096

@pengen_stw

@OdPete

@zenaPSD

Richard Zhang

@rzhang88

2 years

@waxpancake @minimaxir @ByFrustrated Very neat trick to tease this out. Reproduced: - - - I cherry-picked from ~8 generations, since #dalle #dalle2 is adding a different set of word(s) for each generation

23

107

1K

Richard Zhang

@rzhang88

4 years

Dear orange picture people: you obviously loaded as BGR by accident. Simply flip the channels and you're good to go 👍 PC: Alexei A. Efros

7

56

580

Richard Zhang

@rzhang88

3 years

@AlexTamkin This is from buggy implementations in standard libraries, **not a fundamental mathematical issue**. For large factors, buggy implementations are essentially equivalent to naive subsampling. The libraries will be fixed...eventually. Attached image from

8

63

504

Richard Zhang

@rzhang88

4 years

Ask any graphics or signal processing person: *not antialiasing* and ignoring Nyquist sampling theorem is a bug pip install antialiased-cnns import antialiased_cnns model = antialiased_cnns.resnet50(pretrained=True) Now your convnet is antialiased 😃

GitHub - adobe/antialiased-cnns: pip install antialiased-cnns to improve stability and accuracy

pip install antialiased-cnns to improve stability and accuracy - adobe/antialiased-cnns

github.com

4

71

442

Richard Zhang

@rzhang88

11 months

Amazing GenAI results don't just come from thin air! They're a reflection of the underlying training data. Can we specifically identify the highly influential data? See our ICCV work on Data Attribution w/ @ShengYuWang6 , @junyanz89 , A. A. Efros: 1/

6

83

329

Richard Zhang

@rzhang88

2 years

I switched groups in my 3rd year, pushed hard but still missed 2 paper deadlines. The whole PhD thing was not looking very good. After deciding not to submit to CVPR that night, my advisor said "let's chat". I thought I was in trouble...1/2

Ashley Ruba, PhD

@ashleyruba_phd

2 years

Overall, I had a really positive PhD experience. Now that I’ve reached 10k followers (!), I’m ready to share my secret: You must choose an advisor who be kind and not toxic for 5+ years. Oh, and you need decide this based on a few hours of interviews. Don’t pick wrong! 1/

105

525

6K

3

8

240

Richard Zhang

@rzhang88

10 months

I was included on MIT @techreview 's Innovators Under 35 list (at a youthful 34.9)! My appreciation to the wonderful collaborators, mentors, and labmates at @AdobeResearch and @berkeley_ai for the past decade 1/

Richard Zhang

His system combines computation with human judgment to outperform other image-generation algorithms.

www.technologyreview.com

Adobe

@Adobe

10 months

"I'm so grateful to my labmates and mentors, who are world-class researchers," shares Senior Research Scientist @rzhang88 on his name being added to @techreview 's prestigious list of top Innovators Under 35 for his achievements in gen AI & image forensics.

0

1

26

18

5

201

Richard Zhang

@rzhang88

3 years

When I started working on this 5 years ago, I often used family photos to test. My grandma (who recently passed) always got a kick out of it. Funnily, this image started propagating into other people's papers/presentations Happy we've improved @Photoshop Colorize! Try it out

Jess Magdefrau

@JessMagdefrau

3 years

Over here crying my eyes out with this new color changing Neural Photoshop filter. I colorized my late grandma and grandpa's wedding photo 6 years ago in college. It took me a total of 5hrs to do so. The image below, took 2 minutes with one click of a button #AdobeMAX #photoshop

18

26

308

2

16

180

Richard Zhang

@rzhang88

2 years

He said, "Richard, you have all the skills to be a great researcher. Don't worry, keep doing what you're doing, and it will all be okay". A few months later, the colorization paper happened. It all ended up okay. My wonderful advisor and labmates saved my career. 2/2

4

1

183

Richard Zhang

@rzhang88

2 years

I'm...IN!!!! #dalle2 "an asian male happy to be granted access to an AI text-to-image program typing in every waking thought he has into it"

2

8

172

Richard Zhang

@rzhang88

3 years

Swapping autoencoder code by Taesung Park is out now! NeurIPS '20. Also w/ @junyanz89 @oliverwang81 @JingwanL @elishechtman Alexei Efros Github: Website: 3 min video:

Swapping Autoencoder for Deep Image Manipulation

http://taesungp.github.io/SwappingAutoencoderarxiv: https://arxiv.org/abs/2007.00653Swapping Autoencoder for Deep Image ManipulationTaesung Park, UC Berkeley...

www.youtube.com

0

39

162

Richard Zhang

@rzhang88

2 years

5 years ago, @junyanz89 taught me how to hotkey Powerpoint alignment keys, and it changed my figure-making life. Happy ECCVing

1

17

148

Richard Zhang

@rzhang88

4 years

Somehow, our ECCV'16 project website got popular and started propagating around . Four years later, I have gotten around to cleaning it up. Hope folks find it helpful: . Originally made by most helpful postdoc ever, @phillip_isola

2

14

140

Richard Zhang

@rzhang88

5 years

Modern convnets ignore the Nyquist sampling criterion, making them unstable. Come see how simple antialiasing can make your net more stable, accurate, and robust! 3pm tomorrow (Wed) in Seaside Ballroom. #ICML2019

2

30

131

Richard Zhang

@rzhang88

3 years

Download the latest @Photoshop to play with Landscape Mixer, based on our Swapping Autoencoder NeurIPS '20 work (first author Taesung Park). A lot of work from @AdobeResearch + Neural Filters teams to get it into production!

Adobe

@Adobe

3 years

How much autumn do you want? Just say when. #AdobeMAX

6

43

382

4

20

115

Richard Zhang

@rzhang88

11 months

Come check out pix2pix-zero at #SIGGRAPH2023 ! Our twist on traditional image translation -- by leveraging pretrained txt2img models, we can define tasks (e.g., cat→dog) on-the-fly! Talk by @GauravTParmar at 2pm today (Mon, 8/7), Petree Hall D Webpage:

1

19

115

Richard Zhang

@rzhang88

2 years

Short on GPUs/time/data/$, but still want to train a GAN? In "Ensembling Off-the-shelf Models for GAN Training", we show GANs can be "Vision-Aided" with pre-trained nets as discriminators. A #CVPR2022 (oral) by @nupurkmr9 , @elishechtman , @junyanz89 Web:

0

16

113

Richard Zhang

@rzhang88

4 years

@jbhuang0604 🤫 Shh Ok...since we're sharing - Thu is the best; you get the whole weekend - Fri is the worst; you don't appear until Sun, when (a) nobody looks (b) it gets wiped to oblivion by busy Mon Summary 10:59am PT Thu -> best spot on best day 11:01am PT Thu -> worst spot on worst day

4

2

111

Richard Zhang

@rzhang88

3 years

@soumithchintala Fortunately, this is a property of libraries, like @PyTorch , not implementing downsampling correctly, not a math/ML issue. This has created some issues in our ecosystem, for example benchmarking GANs . Perhaps we can reexamine + fix 🙂!

5

13

102

Richard Zhang

@rzhang88

4 years

#ECCV2020 Somebody please enter the schedules into a Google calendar. My PhD didn't teach me time zone calculus okay 😭

8

3

103

Richard Zhang

@rzhang88

5 months

Complaining on the internet: I request @openreviewnet use normal reviewer numbering (R1, R2, R3) instead of random hashes (8FhW, tCqP, mJaY). Cross-referencing uninterpretable characters slows down CVPR rebuttals/meta-reviewing, with no discernible benefit, as far as I can tell

5

4

100

Richard Zhang

@rzhang88

4 years

I think being an AC helped me write more useful reviews. The CVPR AC training video will help you better understand the process, whether you're an author and/or reviewer:

2

8

92

Richard Zhang

@rzhang88

2 years

I don't think we should use "e", like 1e-5 for 10^-5, in papers. My understanding is that it's for calculators limited to 8-segment displays, and it's no longer 1997.

5

1

85

Richard Zhang

@rzhang88

4 years

Come see the 34 @AdobeResearch papers at #ECCV2020 this week! Nearly all papers are results of student internships or university collaborations

Adobe Research

@AdobeResearch

4 years

At this week's #ECCV2020 , @AdobeResearch is presenting new work on research topics ranging from shape estimation to image synthesis, image relighting, geometry processing, facial synthesis, and many more! #ComputerVision

0

22

49

2

5

76

Richard Zhang

@rzhang88

4 years

I will talk about "Style and Structure Disentanglement for Image Manipulation" @ AIM workshop #ECCV2020 Hope to see - West coast insomniacs: 1:30am PT - East coast early risers: 4:30am ET - folks from non-American continents: 930 UTC+1

1

10

74

Richard Zhang

@rzhang88

3 years

Rip the image from the pdf & zoom in. We see clear stippling. Ironically, our computers (or eyesight) antialiases it, so we don't see it when zoomed out or at low-res Hope this is good motivation to learn signal processing 😀

2

12

70

Richard Zhang

@rzhang88

4 years

You don't want a random dog. You want *your* dog. Come see how to find and edit it in BigGAN! Transforming and Projecting Images into Class-conditional Generative Networks w/ Huh (+his dog) @junyanz89 @AaronHertzmann #ECCV2020 @ 4pm PT/7pm ET/00 UTC+1

0

11

72

Richard Zhang

@rzhang88

2 years

First in-person talk in 2.5 years! We'll discuss making generative models faster+higher resolution. So see you at the NTIRE (1:30pm Rm 218) and AICC (4:10pm, Rm 208-210) #CVPR22 workshops. Come see some 9MPix morphs! Can't see that properly over Zoom or Twitter

1

5

69

Richard Zhang

@rzhang88

4 years

Top 87% of NeurIPS submissions 🎉🎉🎉 Feelz good

0

71

Richard Zhang

@rzhang88

4 years

Come visit our poster; we are lonely! We will discuss how contrastive learning can help as a structured loss for image translation problems Webpage: #ECCV2020 page:

1

6

71

Richard Zhang

@rzhang88

9 months

Soon, will we be able to tell synthetic GenAI imagery from real? In "Online Detection of AI-Generated Images", we replay history, studying if leveraging today's models can help detect tmrw's unseen models Come check out our #ICCV2023 Workshop talk, tmrw Oct 2, 8:50am in Rm W07!

5

7

59

Richard Zhang

@rzhang88

11 months

@hitRECordJoe @washingtonpost It's really difficult, but we're pushing to figure out AI attribution, @hitRECordJoe ! Work led by @ShengYuWang6 (PhD student at CMU) w/ @junyanz89 & A.A. Efros

4

9

54

Richard Zhang

@rzhang88

4 years

@jon_barron Yea I found it here: @inproceedings {barron2021general, title={A general and adaptive decaying learning schedule}, author={Barron, Jonathan T}, booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition}, year={2021} }

2

0

57

Richard Zhang

@rzhang88

2 years

Come check out our #CVPR posters tmrw (Thu) morning, 10-12:30! - 4a: Ensembling Off-the-Shelf Models for GAN Training (*also an oral at 8:30*) - 76a: Spatially-Adaptive Multilayer Selection for GAN Inversion & Editing - 77a: On Aliased Resizing & Surprising Subtleties in GAN Eval

1

8

53

Richard Zhang

@rzhang88

4 years

We trained an audio distance metric on human perceptual judgments. To my vision friends: lpips for audio w/ P. Manocha, A. Finkelstein, N. J. Bryan, G. J. Mysore, Z. Jin @interspeech20 "pip install dpam" to try it! Talk: Github:

0

5

52

Richard Zhang

@rzhang88

1 year

@_akhaliq Scaling projects to 22 Billion authors

1

0

49

Richard Zhang

@rzhang88

4 years

Or if you want an AI solution (that works worse), convert to grayscale and colorize. I just converted our 2016-17 models from caffe to @PyTorch yesterday! python demo_release.py -i [[image_path.jpg]]

6

5

47

Richard Zhang

@rzhang88

4 years

Also see this #BMVC2020 oral Delving Deeper into Anti-Aliasing in ConvNets Zou, Xiao, Yu, Lee They show antialiasing also helps instance and semantic segmentation, in addition to classification accuracy

Delving Deeper into Anti-aliasing in ConvNets [oral presentation in...

Project Page: https://maureenzou.github.io/ddac/

www.youtube.com

0

12

44

Richard Zhang

@rzhang88

4 years

Realizing that cleaning old code is probably more useful than writing new (not clean) code pip install lpips import lpips loss_fn = lpips.LPIPS() loss = loss_fn(img0, img1)

GitHub - richzhang/PerceptualSimilarity: LPIPS metric. pip install lpips

LPIPS metric. pip install lpips. Contribute to richzhang/PerceptualSimilarity development by creating an account on GitHub.

github.com

Oliver Wang

@oliver_wang2

4 years

Do you need a perceptual patch similarity metric? If so, you need LPIPS! Now in a convenient to use pip installable package. Nice work @rzhang88

1

17

85

3

11

46

Richard Zhang

@rzhang88

7 months

Shared some experiences about working in GenAI at @berkeley_ai + @AdobeResearch , including - responding to changes in the past decade - slowing down to speed up - coming from a different field Check out @techreview TR35 Festival!

MIT TR35 Festival. Panel on "Accelerating Ideas" (edited)

This is a part of the MIT TR35 Festival, recorded on Dec 6th, 2023: https://event.technologyreview.com/tr35festival-2023/agenda. The description of the MIT T...

www.youtube.com

1

5

45

Richard Zhang

@rzhang88

11 months

Submit 8-page @ICCVConference paper --> Bloats to 9-page arxiv --> Squish back down to 8 pages --> Receive instructions saying camera ready is 9 pages Doh 😂

3

2

46

Richard Zhang

@rzhang88

3 years

Unconditional GANs are now used in the real-world! E.g., @Photoshop Neural Filters. However, they can be slow: StyleGAN2 is 36x more MACs than ResNet50. Our AnycostGAN enables fast previews at lower computational budgets. From @jilin_14 (awesome summer intern) et al at #CVPR2021

Ji Lin

@jilin_14

3 years

Try out the demo and Colab of Anycost GAN: . Our method provides consistent outputs at various computational budgets, paving the way for interactive image synthesis and editing. (w/ @rzhang88 , Frieder Ganz, @SongHan_MIT , @junyanz89 ) (1/2)

1

20

101

1

4

45

Richard Zhang

@rzhang88

3 years

Number of times I've almost revealed myself by signing "Richard" out of muscle memory in CVPR discussions: lots

1

0

42

Richard Zhang

@rzhang88

4 years

@jaakkolehtinen Bilinear downsampling with F.interpolate in PyTorch (Top) Directly downsampling the original by 2x, 4x, ... 32x aliases (Bottom) Recursively applying 2x downsampling looks okay They should be equivalent, but it seems that 2x is safe and other factors are not

4

2

44

Richard Zhang

@rzhang88

4 years

Guest appearance in Yagiz Aksoy+Richard Zhang's ( @richardzhangsfu ) Intro to Graphics: Antialiasing is common practice in classic graphics+vision, but not in modern convnets... yet: pip install antialiased-cnns Also, the course material is excellent!

Sampling & Aliasing (with Richard Zhang from Adobe) - 6 - Introduct...

Find the course website here: http://yaksoy.github.io/introvc/Textbook: http://szeliski.org/Book/Richard Zhang from Adobe Research is our guest for this lect...

www.youtube.com

1

5

41

Richard Zhang

@rzhang88

3 years

Enhance Super Resolution is out! Model development lead by Michaël Gharbi

2

4

39

Richard Zhang

@rzhang88

2 years

Mode collapse -- 4 Richard Zhang's ( @richardzhangsfu , @ryz_nz , @QiuyiRichardZ ) publishing regularly in ML now... 3 of which were overlapping at Berkeley at one point

2

0

38

Richard Zhang

@rzhang88

3 years

Oh yes, I like CLIP very much (I was 31 in this picture)

4

0

35

Richard Zhang

@rzhang88

7 months

I'll be speaking at the #NeurIPS2023 ML for Creativity & Design Workshop today at 1:30pm CT (11:30am PT). The talk will be about "Incentivizing Opt-in & Enabling Opt-out for Text-to-Image Models". Come check it out!

0

4

37

Richard Zhang

@rzhang88

4 years

It seems like NeRFs have caught on even faster than GANs, which took a year or so before the explosion of followups

Frank Dellaert

@fdellaert

4 years

2020 was the year in which *neural volume rendering* exploded onto the scene, triggered by the impressive NeRF paper by Mildenhall et al. I wrote a post as a way of getting up to speed in a fascinating and very young field and share my journey with you:

13

214

952

0

1

36

Richard Zhang

@rzhang88

4 years

Applications to @AdobeResearch 2021 fellowship open on 11/12! Join us for an informational webinar on 11/11: I have collaborated with 2020 winners, Taesung Park and @TamarRottShaham , and won back in 2017, which financed my Honda Civic purchase

Adobe Research Fellowship Webinar

Click to learn more!

adoberesearchfellowshipwebinar.splashthat.com

Adobe Research

@AdobeResearch

4 years

To nurture the next generation of computer scientists, @AdobeResearch annually awards fellowships to PhD students. Learn about 2020 fellows / interns, and watch for 2021 fellowship applications, coming soon!

1

4

30

0

7

35

Richard Zhang

@rzhang88

2 years

Totally unscientific poll -- after attending @CVPR , did you test positive for covid? [1/2]

Yes

317

No

609

6

16

34

Richard Zhang

@rzhang88

4 years

Will your CVPR 202X synthesis method be detectable? If it uses a convnet, we think so! Give it a try: . I recently ran it on 3 submissions. #CVPR2020 Oral: Poster today @4 -6pm, tmrw @4 -6am PT w/ Wang @oliverwang81 @andrewhowens Efros

CNN-generated images are surprisingly easy to spot... for now

Project page: https://peterwang512.github.io/CNNDetection/

www.youtube.com

1

5

33

Richard Zhang

@rzhang88

2 years

Agreed, paper & presentation are very different mediums! I redraw plots and build them, step-by-step. I ALWAYS spend ~5-10x longer than I expect 😂 I also learned from @CarlDoersch to manually animate bar plots in ppt

Dissertation: Image Synthesis for Self-Supervised Representation...

This is my dissertation talk, recorded on April 2018. The talk outline is below:[0] 0:00 Introduction by my advisor, Alexei A. Efros[1] 3:02 Colorful Image C...

www.youtube.com

Jia-Bin Huang

@jbhuang0604

2 years

How to present a line plot? Line plots are effective for describing the relationship between two variables of interests. Unfortunately, most junior students would simply copy&paste the figure from the paper in their talk and cause much confusion. 😕 Let's break it down ... 🧵

6

113

577

1

3

30

Richard Zhang

@rzhang88

3 years

#CVPR2021 in 15min! (10382) Ensembling With Deep Generative Views; @lucyrchai et al (2535) Anycost GANs for Interactive Image Synthesis and Editing; @jilin_14 et al (1509) Spatially-Adaptive Pixelwise Networks for Fast Image Translation; @TamarRottShaham et al See you there!

1

2

27

Richard Zhang

@rzhang88

4 years

I spy with my little eye, some former collaborators 😎

CMU Center for Perceptual Computing and Learning

@roboVisionCMU

4 years

We are excited to have many new PhD students joining our research community this year! Looking forward to all the discussion, collaboration and fun we'll have together.

0

4

33

2

0

27

Richard Zhang

@rzhang88

3 years

Fast image translation (4x-18x faster) using 2 streams: a shallow pixel-wise net at full-res, using parameters predicted at low-res @TamarRottShaham , Gharbi, myself, @elishechtman , Michaeli #CVPR2021 Website: Paper:

Spatially-Adaptive Pixelwise Networks for Fast Image Translation

We introduce a new generator architecture, aimed at fast and efficient high-resolution image-to-image translation. We design the generator to be an extremely lightweight function of the...

arxiv.org

Tamar Rott Shaham

@TamarRottShaham

3 years

ASAPNet was accepted to #CVPR2021 ! Extremely fast image to image translation (18x faster than baselines) with hyper network and implicit functions. With Michael Gharbi, @rzhang88 , @elishechtman and Tomer Michaeli. project: abs:

4

23

140

0

4

27

Richard Zhang

@rzhang88

11 months

We can try attributing synthetic Stable Diffusion images to the LAION training set. But note that there's still a generalization gap from customization → full attribution. We also haven't tackled compositionality. In other words, there's a LOT left to do! 4/

1

11

27

Richard Zhang

@rzhang88

4 years

I'll be speaking at 3pm PT on Analyzing CNN Artifacts in Discriminative and Generative Models: Come check out our just accepted CVPR paper: "CNN-generated images are surprisingly easy to spot...for now" (w/ S.Y. Wang, O. Wang, @andrewhowens , A. A. Efros)!

CVPR AC Meeting

https://docs.google.com/spreadsheets/d/1sEPTaUhhS1Aq4XN7wLMGF0TsVD1FqWQWWa6QbQU-twY/edit#gid=1138630057

www.youtube.com

Greg Mori

@greg_mori

4 years

#CVPR2020 Area Chair workshop features talks from top researchers in #computervision . Happening now, can watch via live streaming link:

1

36

78

1

3

25

Richard Zhang

@rzhang88

11 months

The use of contributor data for GenAI systems is top of mind, with real-world implications. Solving attribution is a critical piece for recognizing contributors See the recent WaPo op-ed from Joseph Gordon Levitt. We even got a shout-out afterwards! /End

Joseph Gordon-Levitt

@hitRECordJoe

11 months

Seems some technologists think it’s impossible to tell which pieces of an AI’s training data influenced which outputs, and therefore which humans would deserve AI Residuals. But some think it’s possible. Here’s a senior research scientist at Adobe. Thanks for this, Richard!

12

37

186

1

25

Richard Zhang

@rzhang88

4 years

Talk is here [32 min] Swapping Autoencoder was accepted to NeurIPS 2020 😬

Style and Structure Disentanglement for Image Manipulation (Aug 2020,...

Talk at Advances in Image Manipulation (AIM) workshop at ECCV 2020. August 2020. https://data.vision.ee.ethz.ch/cvl/aim20/[0:00] Contrastive Learning for Unp...

www.youtube.com

1

3

23

Richard Zhang

@rzhang88

3 years

Adobe Super-resolution was recognized by Time as one of the top inventions in 2021! Try it out in @Lightroom ! Great collaboration between @Lightroom and @AdobeResearch , led by @m_gharbi

Adobe Super Resolution: The 100 Best Inventions of 2021

Find out why Adobe Super Resolution made this year's list

time.com

2

8

23

Richard Zhang

@rzhang88

11 months

Disentangling the contribution from >>100M images & obtaining ground truth attribution is super hard! So we take an initial step. By tuning models towards exemplars using "customization" methods, we can create ground truth training-synthetic pairs, by construction. 2/

1

4

21

Richard Zhang

@rzhang88

1 year

Come to our GigaGAN poster 181 at #CVPR23 #CVPR ! Now to 12:30 @minguk_kang and Taesung presenting, with @junyanz89 and I coming in and out

Minguk_Kang

@minguk_kang

1 year

(1/2) Can GANs be aggressively scaled up? We tried! We present GigaGAN for text-to-image synthesis. It inherits things we like – a disentangled latent space and fast run-time. GigaGAN can generate 512px images in 0.13 sec and produce 4K images.

5

28

174

0

21

Richard Zhang

@rzhang88

2 years

More aspirational #dalle #dalle2

1

0

19

Richard Zhang

@rzhang88

5 years

We detect when a face has been warped by Photoshop and even try to "undo" it. Come see "Detecting Photoshopped Faces by Scripting Photoshop" tomorrow (Fri), 3:30-6pm, Poster #93 . #iccv2019 with S.Y. Wang, A. Owens, O. Wang, A. A. Efros.

0

5

17

Richard Zhang

@rzhang88

11 months

We offer this as an Attribution by Customization (AbC) benchmark, w/ >18,000 models and >4M images. This lends itself naturally to a statement for attribution — given a synthetic image, what % chance is a given image the exemplar? See paper for more dataset & method details 3/

1

2

18

Richard Zhang

@rzhang88

3 years

It took an election cycle, but our Easter egg has been spotted @phillip_isola ! Thanks @BarackObama

Ceyda Cinarel

@ceyda_cinarel

3 years

I have been digitizing my annotations on some paper printouts📜& spotted a couple of gems 🧐a president in the acknowledgements 😎a meme in the citations... so if anyone ever tells you your paper is informal due to something similar just👀

1

0

5

1

2

17

Richard Zhang

@rzhang88

4 years

@junyanz89 First, we solve AGI. Then AGI solves downsampling

0

17

Richard Zhang

@rzhang88

1 year

How much editing is done on magazines? This is running our Photoshop Unwarping forensics tool, from back in 2019! I did a demo here (onstage with @mulaney !): Paper: w/ @ShengYuWang6 @oliver_wang2 @andrewhowens , Alexei A. Efros

#ProjectAboutFace: Adobe MAX 2019 (Sneak Peek) | Adobe Creative Cloud

Identify and undo the Face-Aware Liquify feature in Photoshop with #ProjectAboutFace, powered by Adobe Sensei. Read more about this year's #AdobeMAX Sneak Pe...

www.youtube.com

PetaPixel

@petapixel

1 year

This AI tool lets you see the Photoshop treatment that celebrities get.

0

5

23

0

1

16

Richard Zhang

@rzhang88

2 years

@akanazawa SO EXCITED

1

0

15

Richard Zhang

@rzhang88

4 years

@cvondrick @dimadamen Are you saying the word multimodal is multimodal?

0

14

Richard Zhang

@rzhang88

3 years

Consider applying for the @AdobeResearch Women-in-Technology Scholarship! Applications here:

Adobe Research

@AdobeResearch

3 years

We asked our 2021 @AdobeResearch Women-in-Technology scholars what truly sparks their curiosity about the fields they are studying. Check out what they shared, and don’t miss the opportunity to apply for the 2022 program! #AdobeWiT #Scholarship

0

4

8

0

2

15

Richard Zhang

@rzhang88

4 years

First day of ballot dropoffs, and @SteveKerr was there!!!!! My favorite growing up in Chicago!! Responsible for my earliest sports memory: . I gawked a bit and then ran away 😅. I will make a better plan for next time. Everybody please #VOTE

1

0

14

Richard Zhang

@rzhang88

4 years

Last #ECCV2020 event for me! I will be speaking about "Detecting Generated Imagery, Deep and Shallow" (...and then also showing an image manipulation algorithm 😬) at the SenseHuman workshop 3:20pm PT / 6:20pm ET / 2320 UTC+1 Click Session 2 Come join!

Xiaolong Wang

@xiaolonw

4 years

I am helping to organize tomorrow's #ECCV2020 Workshop (Aug 28) on Sensing, Understanding and Synthesizing Humans tomorrow: Our Speakers include: - @MattNiessner , Pietro Perona, @gan_chuang , @HaoLi81 Kristen Grauman, @rzhang88

0

5

27

0

1

14

Richard Zhang

@rzhang88

3 years

@docmilanfar Worse, "invariance" is overloaded 🙄 - shift(F(x))==F(shift(x)) is "invariance" in signal processing and "equivariance" in ML - F(x)==F(shift(x)) is "invariance" in ML Signal processing came first, but for communication, I just use ML nomenclature when in ML environments

0

2

14

Richard Zhang

@rzhang88

4 years

@roydanroy

0

14

Richard Zhang

@rzhang88

9 months

@nupurkmr9 Poster happening now! Come by and say hi! #ICCV2023

0

14

Richard Zhang

@rzhang88

4 years

EFROS IS HERE

1

0

14

Richard Zhang

@rzhang88

3 years

Did Reddit do this?

1

0

14

Richard Zhang

@rzhang88

4 years

This is great; I would have been super happy to have this resource when I was job hunting

Devi Parikh

@deviparikh

4 years

Introducing AI Paygrades ()! Statistics of industry offers for AI jobs. The goal is to reduce information assymetry so candidates can make informed decisions and negotiate better. Submit your information and spread the word! With @abhshkdz .

42

234

1K

0

2

14

Richard Zhang

@rzhang88

9 months

@ShengYuWang6 This is now! Come on by 😀 #ICCV2023

0

14

Richard Zhang

@rzhang88

4 years

@jbhuang0604 I like a bit of margin, but some students enjoy playing...this most dangerous game:

1

0

12

Richard Zhang

@rzhang88

4 years

@Jimantha Maybe it should be called ABitTooOpenReview

0

13

Richard Zhang

@rzhang88

5 years

Had a fun time presenting at Adobe MAX! #ProjectAboutFace

Adobe MAX

@adobemax

5 years

#ProjectAboutFace knows! This new technology from #Adobe can detect facial manipulation made by the Face-Aware Liquify Tool in #Photoshop .

2

5

26

1

0

13

Richard Zhang

@rzhang88

3 years

@ak92501 Dear @united , please give @ak92501 some wifi credit!

0

12

Richard Zhang

@rzhang88

5 years

Our LPIPS CVPR18 evaluated how well deep activations reflect human perceptual similarity judgments. We did not investigate issues when backpropping it. Check out this paper from @jaakkolehtinen on how ensembling helps prevent LPIPS from being "attacked"

Jaakko Lehtinen

@jaakkolehtinen

5 years

E-LPIPS: Adversarial attacks on neural image similarity metrics, and how to fix them by random ensembling. The resulting geometric properties are perhaps my favorite research finding so far! Paper: Code: @AaltoCS @FCAI_fi @NvidiaAI

1

8

27

0

1

11

Richard Zhang

@rzhang88

3 years

@jbhuang0604 PyPi tracks download stats, which is fun too. Someone clearly had some last-minute ICCV experiments to run here...

0

10

Richard Zhang

@rzhang88

3 years

Anycost GANs for Interactive Image Synthesis and Editing. by @jilin_14 , myself, Frieder Ganz, @SongHan_MIT , @junyanz89 . To appear in CVPR 2021. Video: Website: Paper: Code:

GitHub - mit-han-lab/anycost-gan: [CVPR 2021] Anycost GANs for Interactive Image Synthesis and...

[CVPR 2021] Anycost GANs for Interactive Image Synthesis and Editing - mit-han-lab/anycost-gan

github.com

0

1

10

Richard Zhang

@rzhang88

10 months

@ducha_aiki @giffmana Take a look at this paper by @billpeeb !

0

1

10

Richard Zhang

@rzhang88

10 months

Here's an overview video (5 min) highlighting our work in perception, generation, and forensics: I'm excited to continue pushing to build a healthy ecosystem around GenAI, making it fair and transparent for consumers, creators, and contributors alike 2/2

MIT TR35 Award, Research Overview

MIT Tech Review, 35 Innovators Under 35 AwardResearch Overview videohttps://www.technologyreview.com/innovator/richard-zhang/Works overviewed(1) Perception- ...

www.youtube.com

2

10

Richard Zhang

@rzhang88

4 years

To color is to understand (V4) @phillip_isola ?

Daniel Yamins

@dyamins

4 years

3/7 We found that deep contrastive embedding models achieve neural prediction accuracy that equals or exceeds that of supervised models, in cortical areas V1, V4, and IT all along the ventral pathway. Best-matching algorithm:

1

5

30

1

10

Richard Zhang

@rzhang88

2 years

@MattNiessner Seems folks will individually determine if x% covid possibility is okay and make a personal judgment call. For me, it's definitely worth it (especially if it's < 100 °F and a direct flight away)! I was wondering how this year's covid-positive folks feel:

Richard Zhang

@rzhang88

2 years

[3/2] Okay, bonus poll! Was getting covid at #CVPR worth it? (Poll up for a few days, in case you want to wait to see how your tests/symptoms develop)

1

5

4

0

10

Richard Zhang

@rzhang88

2 years

@AlexTamkin There's an antialias flag now @PyTorch 1.11. Cool! It's set to false by default. So please use it and set it to true!

0

2

9

Richard Zhang

@rzhang88

4 years

@Ar_Douillard I once had a heuristically set hyperparameter of T=1/6. Unfortunately, after release, I found a silly bug in my code (log base 10 in one place, e^x in another). I had to absorb it into the constant to create T=ln(10)/6 🤦‍♂️🤦‍♂️🤦‍♂️

0

9

Richard Zhang

@rzhang88

4 years

Yea! Interns!

1

9

Richard Zhang

@rzhang88

4 years

@ak92501 by @TamarRottShaham , Michael Gharbi, myself, @elishechtman , Tomer Michaeli