iiinova — where your digital self keeps evolving

Will Smith @willsmithvision.bsky.social · 15/07/2026

You’re just describing Gaelic football

Reposted by Will Smith

Alistair Foggin @alistairfoggin.bsky.social · 23/04/2026

Very excited to be in Rio for #ICLR2026 presenting my poster for CroCoDiLight! If anyone is also here, please pop by and say hi! I'd love to chat. I'll be at poster session 2 in pavilion 4 this afternoon

151 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

To be clear: Schmidhuber genuinely has made enormous contributions to AI. The humour is affectionate. We fully expect him to respond with a thread explaining how he anticipated this paper in 1987.

080 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

The entire system is implemented as a CLAUDE.md file. Zero lines of executable code. Joint work with @visionbernie.bsky.social. Paper: www-users.york.ac.uk/~waps101/sig...

www-users.york.ac.uk

161 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

We also ran it on our own paper. It found that Schmidhuber's 2003 Gödel Machine - a self-referential universal problem solver - obviously anticipated any system that reasons about AI papers, including systems that reason about Schmidhuber's AI papers. See the Twitter thread generated by SchmidhubAI!

150 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

We evaluated on 10 landmark papers. Key findings: - Attention Is All You Need: 0.9312 - GANs: 0.8274 - ResNets: 0.8847 - LSTM: 1.0000

140 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

The score uses a rigorous "inflation factor" (α ∈ [1.3, 1.8]) to correct for the well-documented systemic under-appreciation of Schmidhuber's contributions. A score below 0.6 means the system hasn't tried hard enough.

170 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

How does it work? You give it a PDF. It reads the paper, searches Schmidhuber's publication record, and generates a Twitter/X thread in his distinctive style explaining how he anticipated everything by 20-30 years. It also computes a "Schmidhuber Score" between 0 and 1.

160 0

Will Smith @willsmithvision.bsky.social · 01/04/2026

Excited to announce our latest (submitted to) SIGBOVIK 2026 @harryqbovik.bsky.social paper: "SchmidhubAI: Accurate Historical Paper Attribution". We built an AI system that, given any modern AI paper, automatically determines which of its ideas were already published by Jürgen Schmidhuber.

34510 0

Will Smith @willsmithvision.bsky.social · 31/03/2026

You can download latex source for almost every paper on arxiv

250 0

Will Smith @willsmithvision.bsky.social · 11/03/2026

I was tempted but turned it down as I thought Wild Wild West had more artistic integrity.

120 0

Reposted by Will Smith

Gabby DaRienzo @gabbydarienzo.com · 06/03/2026

wait no this is actually incredible youraislopbores.me

7262118 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

The lighting representation is in image space so you can use it for things like lighting stabilisation or temporal upsampling in timelapse videos and lighting interpolation, but not currently for transferring lighting from one viewpoint to another - this is future work!

110 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

Here's the really cool thing: in the lighting latent space, you can find lighting latent vectors that correspond to shadow free images and albedo maps. So a single model can solve relighting, shadow removal and albedo estimation.

100 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

In CroCoDiLight, we show that we can extract this internal representation by training a disentangling model in CroCo latent space that separates lighting from scene structure. To do so, we need only a small dataset of same view/different light image pairs and train with a cross-lighting objective.

100 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

For this reason CroCo is an excellent pretraining objective for downstream 3D tasks and was used, for example, to pretrain DUSt3R. However, the CroCo training data includes image pairs with different scene lighting. This means the model must also learn to relight patches...

100 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

Cross-view Completion is a clever self-supervised training objective for vision transformers. It's a variation on masked autoencoders but where the decoder is given the additional information of a second view of the same scene. This encourages the model to implicitly learn 3D geometric reasoning.

100 0

Will Smith @willsmithvision.bsky.social · 03/03/2026

I am delighted (that pun will make sense in a second) that @alistairfoggin.bsky.social's first paper, CroCoDiLight, has been accepted to ICLR. The idea came from a group discussion on the CroCo paper from @naverlabseurope.bsky.social and realising it might implicitly already understand relighting.

191 0

Will Smith @willsmithvision.bsky.social · 23/02/2026

The LLM obsession with em-dashes has created a weird sort of paradox. I see so much AI generated content that I now see how I should have been using em-dashes all along. But if I start using them, everyone will assume what I've written is AI generated.

130 0

Reposted by Will Smith

Paul @paulpw.bsky.social · 22/01/2026

Excited to share our Paper - VENI: Variational Encoder for Natural Illumination 🌐 🔗 Project page: paul-pw.github.io/veni/ 📄 Paper: arxiv.org/abs/2601.14079 👩‍💻 Code: github.com/paul-pw/veni 🧵 (1/5)

Image comparing the uniqueness of VENI to the Uniqueness of RENI++ by showing an interpolation between two optimized images and showing the "Reconstruction Cosistency": Image error vs Latent error of the two models.

132 0

Will Smith @willsmithvision.bsky.social · 12/12/2025

We need a better name than that. Maybe host it in Denmark and have CVP-Aarhus. Or in a British seaside town and have Southend-on-Sea-VPR.

170 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

Here's the amazing comic-style poster and you can find out more and get the dataset, code etc at our project page: finlay-hudson.github.io/tapvid360/

030 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

It's a really hard task though and there's still huge room for improvement. We believe it might be a promising pretraining objective to help models understand the 3D world before finetuning for some downstream 3D task.

110 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

So we build our dataset by tracking points on 360 videos and then rendering narrow field of view perspective videos with ground truth point tracks outside the field of view. Making some minimal modifications to CoTracker3 and training on this data yields baseline results.

110 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

To solve the task, a model must implicitly reason about the 3D scene, camera ego motion and how dynamic scene contents will continue to evolve outside the field of view. Yet, supervising this task requires only directions, not 3D information, and this is readily available in the form of 360 video.

110 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

The goal of the TAPVid-360 task is to track any point through a video, even when it leaves the field of view. This means that instead of a pixel coordinate or image space representation, you need to reason about direction to the point, e.g. by predicting a 3D unit vector in camera coordinates.

110 0

Will Smith @willsmithvision.bsky.social · 03/12/2025

My students @fhudson.bsky.social and @jadgardner.bsky.social are presenting TAPVid-360 at NeurIPS this week. We introduce an interesting new problem, a benchmark dataset and a baseline adaptation of an existing TAP model for our task. More importantly, they've also created a genre-defining poster...

140 0

Will Smith @willsmithvision.bsky.social · 11/05/2025

Or maybe in the style of Dr Seuss. Or Shakespeare. If so, did it work? Asking for a friend...

210 0

Will Smith @willsmithvision.bsky.social · 11/05/2025

Has anyone ever tried a very non-standard tone for a rebuttal? I'm thinking something like "Hey reviewers! Sit back, relax and let me convince you that you actually want to accept this paper..." or "You wouldn't let a little thing like that stop you accepting the paper would you? WOULD YOU?!!!"

220 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

The date is just an unfortunate coincidence. This is a genuine SIGBOVIK submission and the full working source code is in our arXiv repository.

020 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

This has been a massively entertaining and challenging side project with @jadgardner.bsky.social and @willrowan.bsky.social over the past year and (subject to rigorous peer review) will be appearing at SIGBOVIK 2025.

140 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

@overleaf.com becomes your cloud compute provider. This should level the playing field within the community as both industry and academia will have to work within the same compute limits.

150 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

NeuRaLaTeX brings many side benefits. Your paper source code is also your method source code. No more "Code coming soon" on github. If the paper is on arxiv, the paper source link *is* the method source code! No need to remember those silly git commands anymore!

130 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

In case your arXiv submission is timing out, we've also implemented checkpointing so you can include your trained model weights as a text file with your paper source.

130 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

In a NeuRaLaTeX paper, when you compile your PDF, a neural network is constructed, trained and evaluated with all results and figures generated dynamically. We've worked hard on efficiency and the NeuRaLaTeX paper only took 48 hours to compile.

130 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

Our neural network library extends the autograd engine to create neurons, linear layers and MLPs. Constructing an MLP and making a forward pass is as easy as this.

120 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

Just like Karpathy's micrograd, NeuRaLaTeX implements backpropagation (reverse-mode autodiff) over a dynamically built DAG. However, while micrograd uses 150 lines of Python, NeuRaLaTeX uses around 1,100 line of pure LaTeX, making it about 700% better.

170 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

Not only that, but LaTeX has elegant and intuitive programming syntax. For example, creating a new Value object only uses the word "expand" four times: \expanded{ \noexpand\pgfoonew\expandafter\noexpand \csname #2\endcsname= new Value(\newdata,{\selfid,\otherid},*,0) }

130 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

Wait, what? Well, LaTeX is itself a Turing-complete programming language. When you "compile" a LaTeX document, really you are executing a program written in LaTeX. This program need not only format your paper contents into a PDF but can also perform useful computation.

130 0

Will Smith @willsmithvision.bsky.social · 01/04/2025

So, we wrote a neural net library entirely in LaTeX...

38415 0

Will Smith @willsmithvision.bsky.social · 07/03/2025

On a similar note: if you have two figures one above the other at the top of one column, move one figure to the top of the other column and you gain the white space between them.

030 0

Reposted by Will Smith

Jon Barron @jonbarron.bsky.social · 18/02/2025

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

210924 0

Will Smith @willsmithvision.bsky.social · 17/01/2025

I must confess I just immediately copy/pasted the turkey eggs question into ChatGPT to get the answer myself. Great question!

020 0

Will Smith @willsmithvision.bsky.social · 14/01/2025

#CVPR2025 Area Chair update: depending on which time zone the review deadline is specified in, we are past or close to the review deadline. Of the 60 reviews needed for my batch, I currently have 52 and they have been coming in quite fast this morning. In general, review standard looks good.

030 0

Reposted by Will Smith

Dmytro Mishkin @ducha-aiki.bsky.social · 02/01/2025

Image matching and ChatGPT - new post in the wide baseline stereo blog. tl;dr: it is good, even feels like human, but not perfect. ducha-aiki.github.io/wide-baselin...

ducha-aiki.github.io

ChatGPT and Image Matching – Wide baseline stereo meets deep learning

Are we done yet?

2348 0

Reposted by Will Smith

Gabriele Berton @berton-gabri.bsky.social · 19/12/2024

This simple pytorch trick will cut in half your GPU memory use / double your batch size (for real). Instead of adding losses and then computing backward, it's better to compute the backward on each loss (which frees the computational graph). Results will be exactly identical

3547 0

Will Smith @willsmithvision.bsky.social · 18/12/2024

Of course, we then put our blog post back through the podcast generator! So now its AI podcast hosts talking about our blog post about themselves talking about our research - how meta is that?! They have some minor existential crises as they discuss that they are themselves LLMs. 4/5

100 0