Abe Davis

Honored to be named a Sloan Fellow! So much gratitude for all of my collaborators and others who have supported me! And, of course, a very special thanks to my amazing research group! I am so fortunate to work with such awesome students! #SloanFellow

@MattNiessner Clear dissemination is part of scientific merit, and what is a paper if not text and visuals? LLMs tend to polish by misrepresenting, which should not be confused with good presentation. Academic slop has mostly just raised the noise floor and made good science harder to do.

Historically, academia used presentation quality as a proxy for scientific merit. Now that AI is eliminating polish overhead, everyone is confused, often stuck in debates whether we should allow LLMs. On the bright side, we are finally forced to evaluate the actual research content rather than extrapolating value from the text and visuals.

@jon_barron Ok, but does this factor in the cost of all those unplayed games sitting in people’s Steam library that were purchased on sale because “surely I’ll have time to play that, someday…”?

I asked Gemini to calculate the cost of gaming over the last 50 years. It grabbed a representative game from each year, found and adjusted its launch price, and divided by completion times from HowLongToBeat. Clear trend from a ~$50/hr luxury to a <$0.80/hr commodity.

AI is powerful, but its biggest impact so far has been to erase *useful* friction: proof-of-work we used to assess candidates, curate content, confirm (human) learning, and make fraud harder to scale. "AI slop" is a cute term for something with potentially terrifying consequences

@IanArawjo It's very hard to raise the bar of a publication, so when new tools makes it easy to meet old standards, you get a lot of obvious, opportunistic papers. Criteria like "was this interesting?" or "did I learn something?" are often tragically under-valued in reviews...

Reviewing CHI papers be like: “The rise of LLMs has revolutionized… Yet, X task has not yet benefited from LLMs… We apply a novel LLM-based solution to X… We conducted a within-subjects user study against a non-LLM baseline, finding our solution makes X faster and easier”

Come see @megatran23 present our work on MeCapture at #SIGGRAPH2025 appy hour! Based on our UIST2024 paper with @ethane_yang and @lique_taylor.

Come see our talk on "Pocket Time-Lapse" at SIGGRAPH today at 4pm in the Image Representation, Editing, & Generation session! West Building, Rooms 118-120. With @zzigakovacic , @madhavaggar and @AbeDavis

Come check out our 3 papers @ #SIGGRAPH2025! @liuxr0831 will talk today in the design session at 9:45, @gfxptr today at the Comp Imaging at session 2:15pm, and @ericmchen1 on Wed at 4pm in the imaging session!

Excited to share our ACM TOG/ SIGGRAPH '25 work on Noise-Coded Illumination (NCI)! We use coded noise to add an invisible watermark to lighting that helps detect fake or manipulated video. With @gfxptr and collaborators @zekun_hao and @SergeBelongie. url: peterfmichael.com/nci/

A long a depressing thread

This week in academia, a not so short🧵... 1. Staff reductions and other cost cutting measures coming to Brown University highereddive.com/news/brown-uni…

@getjonwithit @honest_math I think it’s very different. It matters what we automate. People too often try to use AI to automate actual reasoning, where there are opportunities to make new observations and innovate. Foregoing that process has a much greater cost than using a calculator.

@honest_math It's not. I love computer algebra systems, but when I see someone who exclusively uses CASs to the point that they can no longer evaluate a simple integral by hand, I have exactly the same reaction. The key question is whether use of the tool makes the person weaker or stronger.

I get especially sad when I see researchers at my university, including very senior professors, excitedly saying "Look! Now I can use <random GenAI tool> to read papers, write papers, derive equations, write code, ..." And it's like... why are you even here, man? (1/2)

@exJumon @jxmnop Haven't read the paper (yet), but there is value in publishing an explanation of how something works, even if commercial software has a closed-source implementation. The purpose of publication is to share knowledge, not just claim credit for doing something first.

Not really sure if that’s really as big of a deal as it’s being made out to be. Clip Studio has a mode called “perceptual” color mixing. Also, the reason colors historically have not blended as expected when working with physical paint is that in digital you are PAINTING WITH LIGHT. This is basic knowledge if you are a digital artist. You are painting with light. Literally. While the basic colors are red, yellow, and blue, digitally they are red, green, and blue….for a reason. Physical paint and light are completely different things.

no AI here, just the coolest paper i've seen in a while

@jon_barron Amazing movie. I dressed as Arthur Frayn’s floating head for Halloween one year in grad school… 😐

@jon_barron @jbhuang0604 I don’t think this means that text is necessarily “more informative than images”. We shouldn’t be surprised that we can learn more from two modalities than one. Text captures different information, and that information is useful, so including it helps.

@jbhuang0604 I think text is just objectively more informative than images. Image-only self-supervision probably works fine, but just takes longer because the training data is less information dense and therefore harder to leverage.

Why is self-supervision in vision still not working? 🤔 When pretraining a transformer on TEXT-only data by predicting the next tokens, we see clear improvement trends as we scale the model, data, and computing. But after trying to pretrain a transformer on IMAGES-only data with contrastive learning (e.g., SimCLR), masked prediction (e.g., MAE), and self-distillation (e.g., Dino), at the end of the day, the most successful vision encoders still have to rely on vision-language paired data (e.g., CLIP. SigLIP). Or have people not yet tried connecting MAE/Dino with LLMs? Why is that? Thoughts?

Come check out our two talks in the vision-based UIs session at UIST! First paper with @megatran23 @ethane_yang and @lique_taylor. Second paper with @RuyuYan00 and @js3623 !

Come check out our work on Personal Time-Lapse at the UIST demo session! We use AR and custom tracking to capture 3D time-Lapse of the body for healthcare applications. With @megatran23 @ethane_yang and @lique_taylor #UIST2024

Come check out our talk on filter guided diffusion at #SIGGRAPH2024 from amazing students Zeqi Gu and Ethan Yang! In the controllable image gen session at 9am!

I will be speaking today at 2pm in the computational cameras and displays workshop at #CVPR (in Arch 204) Come check it out! csiplab.github.io/ccd2024/

@PandaAshwinee Far from getting rid of rebuttals, I think we should add an optional “pre-buttal” where you can preempt flawed criticism you expect from lazy reviewers with simple explanations that don’t warrant space in the actual paper…

@PandaAshwinee It’s often the reviewers that make the rebuttal necessary, not the paper. As the quality of reviews seems to be getting worse, I think rebuttals are becoming more important, if anything.

as an author and reviewer, please, just remove rebuttals. they are a waste of everyone's time. i have no interest in reading a paper that required a rebuttal to get accepted, nor do i have any interest in my borderline paper get accepted - just reject it and let me improve it.