matroid

I guess it would be a good experiment these days, to see how Nano Banana or OpenAI's latest image to image models will be at transferring lighting while preserving the original style domain, like you suggest.

Thanks for the thoughts :)

That's a good question!

We made the training data for this using a physically based renderer (i.e, Cycles in Blender).

For non-photorealistic art, there are way too many ways in which artists communicate lighting (even in sketch, which is different from pixel-art / video game, there are at least a couple of ways (a) solid shading (b) hatching).

I'm not sure how to extend our methodology (synthetic data in 3D modeling software + train model) to these settings.

Ooh, I didn't know about Myst-style games. I'll definitely check them out!

Hello, I have this project where you can move lights in 3D inside a photograph.

This is the project page and has some cool demos (atleast I think!). Do check them out and let me know what you all think!

Thanks. I'll link it in the first line in the README. I think the interlocking-free part can pack cups like you suggest. They propose a flood fill algorithm which computes all the reachable places for the voxelized shape. It doesn't put assumptions on convexity. I think it would be a great example to try it out on though.

A while back, I implemented a paper that had showed up on HN for a course project (Dense, Interlocking-Free and Scalable Spectral Packing of Generic 3D Objects).

Over the holidays, I cleaned up the implementation (with the help of Claude Code, although this is not an advertisement for it) and released it on GitHub.

If anyone needs fast 3D packing in python, do give this a shot. Hopefully I have attributed all the code/ideas I have used from elsewhere properly (if not, please feel free to let me know).

Hey everyone, I wanted to share my friend's work on Single Portrait Photograph to 3D Head Model. He has a Huggingface demo that you can play with!

Thanks Zie for the message. I'm sorry to hear about your "interpreter" encounter :(

I do think these problems are much, much worse for ISL as you rightly noted.

I think I should have been careful when I said "solve" in my post. But that really came from a place of optimism/excitement.

Also, in India, many hearing-impaired people know only ISL.

That is correct. We want to translate between English and ISL. English, because it is by and large the language of the Web and I think we should try to connect ISL to it rather than Indian Languages.

From my understanding, they are quite dissimilar. A person who knows ISL will not understand ASL, for example.

Thanks for the feedback. You raise great points and this was the reason why we wrote this post, so that we can hear from people where the actual problem lies.

On a related note, this sort of explains why our model is struggling to fit on 500 hours of our current dataset (even on the training set). Even so, the current state of automatic translation for Indian Sign Language is that, in-the-wild, even individual words cannot be detected very well. We hope that what we are building might at least improve the state-of-the-art there.

> It's more of a bad and broken transliteration that if you struggle to think about you can parse out and understand.

Can you elaborate a bit more on this. Do you think if we make a system for bad/broken transliteration and funnel it through ChatGPT, it might give meaningful results? That is ChatGPT might be able to correct for errors as it is a strong language model.

Hello everyone, we are trying to make a large dataset for Sign Language translation, inspired by BSL-1K [1]. As part of cleaning our collected videos, we use a nice technique for aggregating heuristic labels [2]. We thought it was interesting enough to share with people on here.

[1] https://www.robots.ox.ac.uk/~vgg/research/bsl1k/

[2] https://github.com/snorkel-team/snorkel

I have never seen 1x H100 available on Lambda Labs. Don't know why though.

This is such a good resource! Thank you!

All that sounds definitely a lot of fun. I was also working on colorization with SD ControlNet recently (https://vrroom.github.io/blog/2024/02/16/interactive-colorin...).

Wow! Thanks for sharing this.

The author here. I would just like to say that this project is definitely work-in-progress and the AI elements often fail miserably.

As amazing as recent AI progress has been, we do overrate it a lot (I'm including myself in that).

The original blog post by Max Halford (https://maxhalford.github.io/blog/comic-book-panel-segmentat...) does exactly that. I love his approach because, unlike mine, it is simple, yet it goes a long way. I'd encourage you to check it out.

Can you explain what you mean by motion comic generation? Sounds interesting!

Hi, can you explain this problem a bit more. I’m a new PhD student and love low-hanging fruit.

Two reasons (may not be well thought out or wrong):

* Flash Attention, an efficient attention module which significantly speeds up training, only works on Ampere GPUs [1]

* Even if I bought a 3090, I would have to get a computer to go with it, along with a PSU and some cooling. Don't know where to start with that.

[1] https://github.com/Dao-AILab/flash-attention/issues/190

Someone with resources will have to train Zero123 [1] with this backbone.

[1] https://zero123.cs.columbia.edu/

It looks like Vector Networks is based on Boris Dalstein's work (https://www.borisdalstein.com/research/phd). He even has a startup (VGC) for Vector Graphic Editing tool based on these concepts. It is pretty cool!

P.S. I have no affiliation with his work, although I did contribute 10$ to his Kickstarter Campaign back in the day.

Don't know if its worth looking at, but you can look at my code: https://github.com/Vrroom/vectorrvnn :)

I wanted functionality where GPU VRAM isn't constantly hogged while I'm serving a PyTorch model, so that I could simultaneously train stuff.

I wanted a solution which was agnostic to the type of the model, with respect to loading and inferring.

So I made this AutoUnloadModel class that unloads the model if it hasn't been used for some period. I used __init_subclass__ to ensure that all the details regarding timers, locks etc are hidden from the subclass.

I found __init_subclass__ very cool for this job, which is the reason I'm sharing this. Thanks!

Thanks a lot for your explanations :)

What is the guiding principle behind using Swiglu instead of Relu? Did the authors decide by simply trying all available non linearities or is there a deeper reason.

Yup, how did you even think of that! Amazing!

Hello Everyone!

I have implemented a query optimization algorithm for PostgreSQL. The novelty here was supposed to be that the query optimization phase could be split up into sub-optimization steps which could be executed on multiple workers parallely. I think there are not many real world use cases for this. I could only think of a contrived setting where one has to do some data-analysis on many (100s) of data sources, where searching for a join plan might be a bottleneck.

Anyway, as you will find, if you visit the repository, that I found it difficult to use threads/fork processes in Postgres in a way that was reliable and safe. I would like people's opinion on whether the use case I described earlier is a real one and if so, can we parallelize join optimization in Postgres?

I trained a neural network to predict visual similarity between objects in vector graphics. Using this, I built a tool for smart selection. The tool, when working correctly, lets you select related elements of your initial selection automatically.

In the url above, I test out two models of visual similarity (one that I developed and one that has been previously developed) to power the tool.

Hello everyone,

I have made a tool, as part of a research project, that tries to make it easier for people to edit Vector Art. In the study/survey posted above, I have implemented two algorithms that try to define what "perceptually-guided" might mean. I want to know which of the two is better and collect data on how people use this tool. To achieve this for each session, I randomly pick an algorithm and use it to drive inference when the tool is being used. I'm aware that neither of the two algorithms are close to perfect but want to know whether this line of work is relevant for creators of Vector Art.