Turn your picture into a Lego character

https://youml.com/recipes/236-photo-to-lego-minifigure

by alexp7 • 3 years ago

18 16 3 years ago

youml.com

carterschonwald• 3 years ago

Seems to be hugged to death with an OpenAI rate limit.

minimaxir3• 3 years ago

That's odd IMO since none of this workflow should be hitting OpenAI.

jeffbee• 3 years ago

Error code: 429 - {'error': {'message': 'Rate limit reached for gpt-4-vision-preview in organization org-eRnbbdbvRysampoMyGNkYZ8A on tokens per min (TPM): Limit 10000, Used 6380, Requested 4151. Please try again in 3.186s. Visit https://platform.openai.com/account/rate-limits to learn more.', 'type': 'tokens', 'param': None, 'code': 'rate_limit_exceeded'}}

alexp7• 3 years ago

You guys broke it! :)

Apparently there is a 7-image per minute hard limit on OpenAI image generation, so come back a bit later.

diegorbaquero• 3 years ago

Why restrict it to only Google accounts?

andy99• 3 years ago

Presumably to harvest sign-in information

spencera• 3 years ago

Ever so slightly harder to create dozens of accounts to get free credits?

AnotherGoodName• 3 years ago

It's a cheap way to prevent abuse.

If you have no log in flow at all you'll be spammed. If you have your own sign in flow how do you detect and stop automated account creation and subsequent spam? Can you beat Google at stopping bot signups?

In the end it's a lightweight barrier to abuse. You can argue they should diversify and not just use Google but hey it's a small project. It's understandable. It's also free to sign up for Google for a real user.

eclo• 3 years ago

didn't work for me :/ got a safety system error

Obscurity4340• 3 years ago

This one right here, officer /s

deekayw0n• 3 years ago

yeah... 429 rate limited requests... and dennnnnnn out of credits :(

x0n• 3 years ago

Hmm, yandex.ru backend? No likey.

nicholasbraker• 3 years ago

Red flag for sure

minimaxir3• 3 years ago

Some notes on how you want to do this yourself:

1. Get atleast ten 1024x1024 images of a style you want (e.g. Lego). You can use SDXL/Midjourney to generate these if you are good at curation.

2. Train a LoRA on SDXL using those images.

3. Construct an image-to-image pipeline with the LoRA'd SDXL and pass along the input image as the base image. You can run for fewer steps (20-30 instead of 50) and/or preprocess the input image (e.g. segmentation).

There is a lot of other parameter tweaking for both training and generation which is where the added value is for websites such as these. But with a service like Replicate the pipeline is relatively cheap (~$0.50 for training, $0.01 for each image generation). And will only get much cheaper once people play around with LoRAs for smaller and faster models like ssd-1m-lcm.

jansan• 3 years ago

Good thing it's free, because it failed but still took my credits.

Turn your picture into a Lego character

Related Stories