01 March 2025

IISc Open Day 2025

Open Day at IISc is one of those blissful times when you can enter a haven of intelligence and creativity while the rest of the world is busy with corruption, childish politics, superstitious idiocy and mindless entertainment. I wish there was such an environment I could have visited every weekend as a child. Anyway, since this time IISc was very crowded and there was a lot of walking to do, I decided to visit only a few projects. Here are some I found interesting:

At the brain research center where they do research on Alzheimers and dementia:

Avoiding metabolic syndrome

Get your vitamin D

Neurogenisis is better in people who keep their brain active

This device can tell if you have peripheral neuropathy by what level of vibration you can detect

Presynaptic neurons

Brain degeneration in dementia

Interesting quiz for kids

This was at the main campus:

Play chess

Have icecream made with liquid nitrogen cooling

Sodium alginate is obtained from seaweed

Sodium alginate forms a gel which can help disperse medicines in the body precisely

Kids punching a non-Newtonian fluid

3D printing metal on metal. This allows printing a magnetic metal onto non magnetic metal too

Vortexes in water can make flow inefficient. This lab experiments with various types of pipe materials to make flow efficient.

A mini seismograph

Using transducers to measure shockwaves created which are reflected back to the surface

Various ways of minimising landslides

Demo of how a grid polymer can keep loose sand in place

Crowd tracking using optical flow and having an early warning detection of stampedes

The people who proposed a suburban railway for Bangalore

Re-generating what a person saw, using fMRI scans

Diffusion techniques are used for regeneration

The model is first trained on known fMRI data generated when a person watches images

Using a specialized model to detect emotions in speech

Methods of improving AI reliability

There were many games based on AI projects. You can win chocolates as prizes.

Using Google's Teachable Machine to train an AI

Use gestures to play space invaders

Take a photo and substitute it in an AI generated image

Building 3D models using depth cameras

Recreating a jar in 3D from few photos. It can't handle many photos due to the small context window of vision transformers

Robots avoiding obstacles live

The obstacle avoidance control panel

Levitating metals via magnets

Large crowd, coconut water at Rs.50, free Bisleri water, toilets and games

As with every year, there were two places where they showed demos of the non-Newtonian fluid and two places where they showed demos of instant ice cream. Interestingly, one of the projects demonstrated how slowly water seeps into the soil, and mentioned that it can take a thousand years for water to seep through and reach the water table below. Of course that's not the only way it reaches.

There was a challenge given to us by some computer science students, to find a flaw in their simple coffee ordering program. The program gives you an account balance of Rs.100. You can either buy a coffee for Rs.35 or buy the special coffee for Rs.10000. The challenge is to find a flaw which allows you to buy the special coffee. I solved it within a minute, because I figured it might be an integer overflow problem. So I typed something like -10000000 at the number of cups to buy, and was able to 'buy' the special coffee because an integer overflow increased my account balance. There were no prizes for solving it, sadly. These guys had actually allocated two hours of time for solving it :) They were surprised I could solve it in one shot, because as per their method, it required multiple attempts at increasing the account balance.

It was nice to see students working on the latest tech in terms of LLM's, but there were other groups that were using LLM's for image processing, but they were not even aware of basic image pre-processing techniques, which could have solved their problems much more easily. They were oddly struggling, saying that they were getting different outputs in different Android versions. I gave various teams many practical solutions which they agreed with and were grateful for.

One researcher mentioned that she detected lead in milk samples in Bangalore. She didn't reveal the brand, but said she tested seven brands, so I guess one or more of those contained lead (which can be harmful to the brain and nervous system).

One of the teams actually asked me for funding, which I found very strange. It is sad that people doing such excellent research are made to struggle for funds, while people spreading superstition obtain funds relatively easily.

There were some interesting games too. One of them required me to hold a 20kg dumbell horizontally with my hands fully extended. I managed to hold it for one minute and eleven seconds. They said it was a good duration, and that the maximum someone had held it was for two minutes. The guy who tried before me could hold it only for 40 seconds.

Most of the food stalls in the main campus appeared to be doing brisk business despite the quality food not being too good. Some of them accepted payments only via UPI. The food stall at the brain research center however, did poor business due to a lack of people visiting. They could've offered food at a lower rate. Somehow, I couldn't trust the quality of food, and even at the IISc canteen, the food was priced too high. For such an event the stalls and canteen really could have offered food at a lower rate. So if you go there, make sure you carry your own food. Also, the gymkhana grounds used for parking, leaves a lot of dust on the vehicle, so public transport would be better. Alternatively, they could have sprinkled water on the grounds periodically, but March is a time of water shortages, so perhaps Open Day could have been held in Jan or Feb. It'd be cooler then too.

Visiting IISc was as usual a pleasure, and if it weren't for the yearly increase in crowds and poor public transport access, I'd love to go there again.

02 February 2025

Free GenAI text to image generators that can be used commercially

I wanted to generate some images using AI, and wanted to find out which ones were completely free to use for commercial purposes. Various such tools obviously had all kinds of tools to customise the image, but for a trial, I tried a simple prompt only.

Disclaimer: These were based on a cursory survey. Please double check the terms of use and any other conditions imposed by all the websites. I am not liable for any errors.

Prompt used

"A comic strip of a hero named izak who suffered chronic eye strain and recovered and is now trying to save the world from the evil villains who trap children and adults with eye strain, spectacles, eye tests, lasik and eye drops by making them addicted to games, TV, smartphones and by making them work long hours in office in front of computers."

Note: the word "trap" was earlier "enslave", but Blogger's keyword filters blocked this from being published and then I had to appeal for a review and then it got published. I've faced similar irritating keyword filter issues with some websites where I tried generating images, which shows an unfortunate reality that even websites that use generative AI, are not using generative AI to evaluate prompts.

Copyrights issue

The copyright for using AI-generated images is a complex legal issue that has not been fully resolved in many jurisdictions. In general, the copyright for an AI-generated image is likely to be owned by the person or organization that created the underlying AI algorithm or data set that was used to generate the image. However, if the AI algorithm was trained on copyrighted images or other materials, then there may be copyright issues related to the use of the AI-generated image. Perhaps projects like H20AI may be free of copyright issues.

The websites below are randomly ordered.

OpenArt

Openart.ai has image to prompt too. Their style palettes are amazing. They have a 7 day free trial with 40 credits. The generated images can be used commercially with credit being mentioned. You can fine tune the model with a model type and uploaded images. They also have various apps for facial expression changes, changing parts of images in various ways, changing styles of objects in images and more. It didn't understand that I wanted a comic strip in Pixar style.

It did a poor job of understanding my prompt of "change the spectacles to rimless" which I did after selecting the inpaint area tool. With better prompts I guess things might be better.

Deep AI

I think this is supposed to be free forever. The image generated was not bad at all, though it didn't understand that I wanted a comic strip. It didn't even need me to login to generate this. They also have an AI video generator and chat. As per the terms of use, the generated images are free of copyright and can be used commercially.

StableDiffusionWeb

This was the first one that created a proper comic strip, but the words were gibberish and the spectacles on the characters was weird. The free plan allows 10 credits per day. Images created through Stable Diffusion Online are fully open source, explicitly falling under the CC0 1.0 Universal Public Domain Dedication. They even have a Prompt Database. It added a watermark though at the bottom right.

Playground

They allow 15 free images every 3 hours and images can be used commercially. Rather than being for general image creation, they allow Logos, Art, Stickers, Wallpapers, Posters, T-shirts, Monograms, E-book Covers, Cards, Invites, Patterns, Mockups, Memes and more. It did a decent job.

Freepik

I liked Freepik. It's a Spanish company whose website is well thought through and intuitive to use. They have a host of tools to use, like retouching, background removal, re-imagining, mockup generator, image editor, upscaling, video generation and more. They have various free limits for stock images, AI tools (20 images per day for generating images), and more limits for other tools. You can use AI-generated images for personal or professional purposes, including commercial projects, provided they comply with applicable laws and respect third-party rights, such as identity, creative ownership, trademarks, and intellectual property. However, it is your responsibility to ensure the images do not infringe any third-party rights and comply with the applicable regulations. Their terms of use are here.

Ideogram

They have website which takes long to load any page initially, but the free plan has 10 credits, where each credit generates 4 images. The results were pretty good. They do not claim any ownership rights in your User Input or User Output, and they do not restrict your ability to use User Output for your own purposes (including for commercial purposes). Ideogram seems to be built by a good team.

Starry AI

They give 5 free credits daily and generated images can be used commercially. They had a reasonably good website too. It allows uploading more photos to train it to a certain style. The generated image was too basic though, and didn't form a cartoon strip.

ComicsMakerAI

They give 100 free credits per month. Generated images can be used commercially. They created a nice process where you can create characters using a prompt, create a scene with another prompt, create a background with another prompt and then there's a pose editing tool to adjust the pose of the character. But when the character didn't conform to the pose I created, I realized that such tools have a long way to go before being usable.

Generate images on your own computer for free

Generate images with your code and HuggingFace's servers

Deep Floyd
Hugging Face Inference API which allows 1000 requests per day. I don't think they store generated images on their servers, but since it's not mentioned anywhere, it's better to check. An alternative that uses cleaner sources of images may be H20AI.

Other text to image generators with limitations

Flux AI

Did a poor job with generating the cartoon hand and the "I" in Izak. There were plenty of other errors in the image. They offered only one free credit. Didn't like it.

Flux AI.art

At least this LLM understood that it was supposed to generate a comic strip. Here you can use Flux.1 by allowing experimentation for free, but they retain the copyright. They also have a table comparing Flux1, midjourney, Dall-E and SD3-ultra.

Stabledifffusion

With an extra "f" in the name, this offers free image generation and seemed to understand that a comic was required. This is generated on a UK server and didn't mention anything about commercial use, so I assume it does not allow commercial use.

Leonardo AI

They provide 150 tokens free everyday, but I found it extremely irritating to use because when I typed the prompt, it showed "Error. Phrase blocked". It didn't even show me what phrase was blocked. I tried altering the prompt using the prompt generator and it still didn't work. The engineers working here seemed careless. I promptly deleted my account.

Midjourney

Subscriptions start at $10 per month, so no free generations.

Dall E

This can be tried in ChatGPT itself, but even this did not generate the image because it apparently didn't align with their content policy. Another dumb implementation of security.

Stability AI

Has only a free trial, but no free credits, so didn't try it.

Rendernet

This website produces amazingly detailed images and videos, but has no free plan.

GetImg

They give 100 free credits free per month, but no commercial usage rights for the free plan.

FluxAI image generator

They give 100 credits, 5 per day, and no commercial usage rights for the free plan.

Bing image generator

Provides 25 credits and does not claim ownership of generated content, but says nothing about commercial use, which means it cannot be used commercially. They did a somewhat good job of generating a comic, but all with superman instead of a unique superhero.

Canva

Canva's dream lab was so far the best. They offer 25 free generations in total and it can be used commercially, but under their terms. So not really a great option for free image generation.

Lexica

Has only paid plans.

Wonder

Has only 5 free credits and very few controls, but good generation capability, although it didn't understand I wanted a comic strip.

Hamster AI

They have a surprisingly high number of free credits. Without signing up itself there are 5 credits. With a signup, there are 250 free credits per month. They advertise themselves as being among the most cost effective AI website and offer more tools for chat, searching the web, etc. Nothing was mentioned about commercial usage of the free images.

Channel AI

They needed an app install on the phone and the free generation was only for limited time, so skipped this.

Photoleap

This appeared to mainly be for photo editing, and had only a free trial, so skipped it.

AnimeBG

This seemed fine tuned for anime images only, so skipped it.

DesktopHut

This was a poorly constructed website with good controls for AI generation. The image generation seems free, but they seem to have restrictions on copyright.

Pictrain

They seem to be particular about copyright, so didn't proceed. This service seems particular about being able to generate art in the style you specify.

Mage Space

Was not to happy with the website and the generation. They initially showed 10 free credits and then with the generation of the first image, it said "You've reached the end".

Scenario AI

They don't have a free plan.

Minimax ai

This turned out to be for free (for the time being) video generation, but it comes with a watermark even though people are allowed to use the videos commercially.

ImagineArt

They offer 1000 free credits for 3 days for video generation. They have image to video and text to video. Once a generation began there didn't seem to be any way to cancel it, which I found silly. The website layout was also poorly thought through. It's for non commercial use only. It generated a 1280x720 sized video of 5 second length which was 6.1MB in size. Poor job on video compression.