ivanmikhnenkov

/

tinydit-256

Like 5

Hey Ivan!
I saw your TinyDiT-256 model and I am amazed by the results and the image quality of the outputs.
I got your model running instantly on my Kaggle Setup and tried some samples:

Prompts:

PROMPTS = [
    "an astronaut riding a brown horse on mars, morning light, dust",
    "the Google logo",
    "two yellow wooden bricks and a yellow cup of coffee",
    "a water bottle standing on the floor",
    "a house in the woods",
    "a cliff coast, crashing waves in the bottom, breathtaking view",
    "a bed room"
]

I wanted to ask you, if you want to join SupraLabs as a new member for pushing our own text-to-image models further.
We did a lot of research before and we never got results like you did 😕😅

For datasets, we would propose almost the same datasets as you used (it's a funny coindicent because I found your model AFTER I did the dataset research and your GitHub-model-README just instantly proved my decisions 🤗😀🤩 - that's so cool! :D), so:

• ~75% cc1m-cleaned (new source of REAL images and good prompts!)
• ~25% FLUX-Reasoning-6M (same dataset that you used!)

And then a total of e.g. 10M images for ~2-4 epochs.

We are learning a lot from your model setup - and I want to cooperate with you to make a model that is even greater!

At SupraLabs, we soon have 24/7 FREE access to a new server we got from a friend: 2x RTX A6000 48GB (96GB via NVLink), 128GB DDR5 RAM, 4TB NVMe SSD 🔥
--> That would be the hardware we'd train the text-to-image model on.

And a question to you: which GPU ("from scratch on one GPU") did you train on?

We'd be happy to work together with you 🤗
Please tell me if you're interested to be our new member for Computer Vision and Media-Generative AI (t2i, t2v, etc...)!