Thread by @atroyn

announcing stable attribution - a tool which lets anyone find the human creators behind a.i generated images.

stableattribution.com

given any image generated by stable diffusion 1.x or 2.x, stable attribution finds the images in the model's training set which most contributed to the generated image

we (chroma) have been in crash development on this since mid-december. once we saw the technical approach was feasible, all our effort went into making this easy to use, scalable, and highly available

the discourse around generative image models, especially the use of creative's work in their training sets, has grown ever more difficult.

i was often in error myself, but things built up so i kept digging

https://twitter.com/atroyn/status/1602384325403631618

many of my assumptions were wrong
- the training data was not public domain, and often not licensed at all
- the models do replicate images in their training sets, and can be made to do so reliably

i started paying attention to what creatives were actually saying.

historically, creatives have been among the first to embrace new technologies. ever since the renaissance, artists have picked up every new tool as it's become available, and used it to make great things.

these people aren't 'luddites'

creatives want credit and compensation for their work. seeing generative image models in action shocked them (as it did everyone), but still more shocking was the fact that their work had contributed to these models without their knowledge or consent

this is completely normal, and should be very familiar to technologists. the first thing after the title of every scientific paper you read is the list of the contributing authors, details of their affiliations, in rough order of contribution.

the responses from technologists were, frankly, embarrassingly midwit - either attacking artists as 'luddites', claiming that the machine 'learns' like a human so artists have no claim to attribution on the derivative works (wrong), and a bunch of other pretty dumb things

here's a particularly embarrassing example, which manages to be simultaneously wrong about
- how image compression works
- how latent diffusion models work
- whether they can reproduce images in their training set
- what a derivative work is, legally

understanding what models are doing based on their training data is core chroma technology.

so instead of wading into the discourse, we decided to try to solve the actual problem - source attribution

i'll post some technical details in another thread, but the short version is we found an approach that would work for stable diffusion - the training set (LAION) is publicly available and the weights are open source, which allowed us to build & test effectively

what we have right now is the first cut of the algorithm - we wanted to get this out fast, while it still mattered.

i dreaded waking up one day in a horror hellscape of dmca 2.0, hence crash development rather than an extended research program.

it can be made to work better.

it *only works correctly* if you upload images from stable diffusion 1.x / 2.x

we cannot stop you uploading any image you want into it, because almost all available hosted versions of stable diffusion have the watermark in the oss repo removed

https://twitter.com/atroyn/status/1622361964419899392

here's what we hope to achieve:

- show that attribution is possible. this lets creators get credit for their work, and opens the door to them being compensated for every generated image their work contributed to / influenced.

besides being the right thing to do, crediting and/or compensating creators for their contributions to generative models creates a virtuous cycle; cutting them out is eating the seed corn for future models

- get ahead of the legal situation before it goes really badly for everyone. generative image models are amazing technology; stable attribution demonstrates they can be made and used in a way which benefits everyone involved

i spoke to law profs about this - the analogy which kept coming up is the vcr. initially basically a piracy machine, it brought to life an enormous content market. had it been banned, creators would have been worse off in the long run.

here's how you can help.

- help us find links to artists' work. we have urls for the source images, but they're often served by a cdn, so we can't link directly to profiles, though we *really* want to.

- get the word out, especially among the creative community. there is a better way.

i’ll be doing a twitter space about this tonight including time for q&a twitter.com/i/spaces/1jMJg…

the purpose of technology is human flourishing.