Diego Porres

Fun weekend project: explore the latent space with your hand! Using SD-XL turbo and some pre-defined prompts, this is a proof of concept, but we plan to do so much more. Stay tuned :)

Super-Resolution has been a widely studied field, yet its metrics haven’t kept pace with the methods. Check out RQI, a new perceptual metric that better aligns with human judgment. RQI will be presented at #CVPR2026 Code coming soon, enabling better SR models!

Start of Day 2 of the @ELLISforEurope PhD school! First, Robert Geirhos from @GoogleDeepMind on his personal top 10 lessons for future researchers. Relevant advice for folks in academia and industry!

@atulit_gaur You’re training end-to-end driving models and val loss is useless

a question to ask in ml interviews for three consecutive epochs you don't see any meaningful decrease in val loss but then on the fourth epoch you do, why?

I almost got held back in 5th grade because I couldn’t focus in class. All day long I could only think about getting home and playing FF7.

I thought the term world models was misused before (> 2 years ago), but nowadays I'd even venture to say that term has lost all meaning

@AlsikkanTV This gave me whiplash to Twitter from the 2010s, thank you

just met someone named Cheddar Larson and she said her sister is a famous actress but she wouldn’t say who

@leothecurious Holy shit

it's refreshing when two different hypotheses i've been excited about get validated in a single paper. tl;dr: convolutional inductive biases in early stages of visual processing, and latent prediction of global semantic features from local spatial context, can both aid in achieving higher sample efficiency on visual tasks.

@phillip_isola This is just Zeno’s paradox for the modern times

As models advance and surpass certain human abilities, “human-level” advances too, as we can use them as tools. So yes a model might do better math/coding/etc than I could have done in 2025. But they still are behind where I could be in 2026! This thought gives me some hope :)

@quantbagel Nice! These are offline metrics though, have you been able to deploy these into a real/sim robot? Note if your inference speed is much higher, it might also perform better (maybe in general, maybe only on tasks that require finer motion).

Robot action models shouldn't need 256 vision tokens per frame. Pi0.5 spends 400M parameters on SigLIP just to see. We replaced it with a 4.4M encoder that outputs 5 tokens — and action quality barely changes. 91x smaller. 51x fewer tokens. 7.3x faster inference.

Introducing @QuiverAI, a new AI lab and product company focused on frontier vector design. We’ve raised an $8.3M seed round led by @a16z, with support from amazing angels and investors. Our first model, Arrow-1.0, generates SVGs from images and text. It’s available now in public beta at app.quiver.ai

@joanrod_ai @QuiverAI @a16z Congrats Joan! It’s been a long trip but I’m certain great things await you and the team!

@jon_barron You can always mess with some weights of the network for it to fail by the right amount

Unfortunately these gifs were time-gated to 2025 because they're contingent on 1) models being bad at reproducing the input, which got fixed, 2) platforms subsidizing many independent generations, and 3) human willingness to do a repetitive manual task, which is now agent-work.

@kamalgupta09 @ashishkr9311 do you need video to action if your model can run at 120 FPS?

The post triggered a lot of 3D vision folks but it is right on money. Had a similar epiphany regarding robot learning ~2 years ago after a long chat with @ashishkr9311. 3D priors may give you short term efficiency gains, but long term going from video to action, and allowing big models to learn their own intermediate representations is the right direction.

@iScienceLuvr Goldstein?

Image Generation with a Sphere Encoder a few-step image generation method by mapping images to a spherical latent space, trained with simple reconstruction+consistency losses

@jordiponsdotme the whole building is wrong, but yes

the map is wrong

@sea_snell lmfao

People who do model merging are like the flat earthers of deep learning

@gabriberton Genuinely think this is one of the best papers and it's underused. The most recent paper, of the few I've seen, that use it is RAP: arxiv.org/abs/2510.04333

This would force the features extracted by the model to be domain agnostic, i.e. the feature contain no info about the domain, making the features more robust on the target domain Cool stuff

A little more info on Domain Adaptation: the task is that you would have a labelled train set of one "source" domain (e.g. daytime images) and an unlabelled set from the test/target domain (e.g. night images). [1/N]

@hengcherkeng @rsasaki0109 literally the Matrix plot lol

@rsasaki0109 It might just as well the agent writes the paper, makes rebuttals, polishes and makes final submission. Then the agent books air tickets and attends the conference and writes reports. The human author did nothing except paying for electricity bills

Paper2Rebuttal RebuttalAgent: AI-Powered Academic Paper Rebuttal Assistant github.com/AutoLab-SAI-SJ… RebuttalAgent is an AI-powered multi-agent system that helps researchers craft high-quality rebuttals for academic paper reviews. The system analyzes reviewer comments, searches relevant literature, generates rebuttal strategies, and produces formal rebuttal letters, all through an interactive human-in-the-loop workflow. Key Features 📄 Automatic Paper Parsing: Converts PDF papers to structured text using Docling 🔍 Issue Extraction: Breaks down reviewer comments into actionable issues with priority levels 📚 Literature Search: Automatically searches arXiv for relevant supporting papers 💡 Strategy Generation: Creates data-driven rebuttal strategies (not sophistry!) ✍️ Rebuttal Writing: Generates formal, conference-ready rebuttal letters 🔄 Human Feedback Loop: Iteratively refine strategies based on author input

@Birchlabs It's just trying to ragebait you

after hours of debugging, got to the bottom of why training was diverging this whole time I was doing gradient ascent

@Algomancer StyleGAN models did this with the disentangled latent space W. From what I've tested, they reach almost the same family of distribution, but this is still to be finished

Question for my Flow Matching / Diffusion pilled friends. I've been doing this for years but never seen it on my feed. (I havn't actively looked for it, so if you know any reference papers, kinda just seemed obvious). I use it for my diffusion/flow matching prior vaes, but it works fine in rectified flows / mean flow / etc recipes where your focused on reducing the number of function evaluations. Do people ever learn the prior/starting distribution? ie where the noise distribution (prior) is learned rather than fixed to N(0, I). (Quick toy example below from some of my adverserial flow matching experiments so you know what i mean). The intuition being that optimal transport cost depends on the choice of source distribution. A learned prior reduces the total transport distance by better aligning with the data geometry. github.com/Algomancer/Adv…