Releases: unconst/ImageSubnet
v0.5.7
v0.5.1
Update v0.5.0 - Image To Image!
In this update, we change the validators scoring mechanism in two parts.
Image To Image Scoring Method
First they will now perform an image2image query on each and every loop. It works like this...
25 T2I queries -> score and select best image -> pick 25 new miners -> query them w/ I2I and update their weights as well
Miners who support i2i will be scored much higher than those that dont as those that dont will essentially lose out on half of the rewards possible.
Requery
In 0.5.0 we add a sqllite database to manage and store recent historical prompts sent to the network. This allows validators to requery miners to ensure they output the same image for a given prompt + seed combination.
In this reward method we do not care about image quality. If a miner output a crappy image for a prompt before, so long as it outputs the same image again this time it receives the full reward. In this way, miners do not compete at this level, they all have the same chance to get the same reward value.
Prompts are held in the database for 48h, after which time they are removed.
Other changes
v0.5.0 also makes some progress on cleaning up the main function, making it more legible and easier to read by 3rd parties. In no way is the file in its final form for code cleanup, but hopefully these changes allow for an easier time by other developers to figure out how the scoring mechanisms actually work.
A generate.py script was also added. This can only be used by validators on their official IP they have broadcasted to the network. However it allows the owners of validators to mass query the network and view results for a set of prompts to see what the true quality of the network is at.
Future Changes
Some issues that have arisen recently are miners that are providing faulty images back to user prompts but not actual validation prompts. This could be due to a number of reasons but the main being that fundamently the validation prompts dont look like user prompts. This allows miners to detect the difference and respond back with poor images to the network. I've decided not to opt towards a black list method for now, and instead have been working on a more long term solution.
v0.6.0 will include a way more advanced prompt generation method. It's quite a bit resource intensive for validators so be prepared to increase your vram on validation servers to a 3090 at least. You may be able to get away with running the smaller version of the model on 4bit in a 12gb card, but it's not advised.
This change was going to be included in this update however its taking quite a bit longer to work out all the kinks than I was expecting, so I'm splitting the releases into two with the next planning to come out a week from now if all goes well.
As always, thank you for the ongoing support and may the τao guide you 💙