6 ms·
so a few days back, I discovered that with the new Flux v1.1 pro image gen model, if you use a prompt like "IMG_XXXX.HEIC", you get some extremely realistic ima
by spaceman_2020 2y ago
so a few days back, I discovered that with the new Flux v1.1 pro image gen model, if you use a prompt like "IMG_XXXX.HEIC", you get some extremely realistic images.
I spent a few hours playing around with it, trying out a ton of combinations. Ended up with a very large library of images and decided to make a quick website to share it.
Some notes and observations:
1. Using a prompt like "IMG_XXXX.HEIC" tends to yield the most realistic images, but most of these tend to be rather mundane images of landscapes, flowers, poorly shot cityscapes
2. Adding "IMG_XXXX.HEIC posted on Snapchat in [year]" yields more realistic, casual images of people. However, a lot of these look like screenshots, complete with the Snapchat UI. Most people also tend to be attractive.
3. Adding a [year] in the prompt yields some interesting images. Like [2017] will yield blurrier images than [2023]. Adding [2021] in the prompt got images with face masks and face shields
4. The prompt "[firstName] [lastName] selfie" gets real-looking selfies of real-looking people. You can use Indian, Hispanic, Chinese, European, American, etc. names and get realistic images of people with these ethnicities. Example: https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/pc5JnhNRUEqu4FMSBbhA4.jpg https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/pc5JnhNRUEq...
5. There is a decently high failure rate. The ~750 images on the site are hand picked. I had to delete around 220 images for not meeting the criteria (not real enough) or being just bizarre
6. If this model is any indication, its soon going to be impossible to tell what's real online
- puttycat 2y agoGiven your technique, how can one verify these aren't just copies of the training data?
- WithinReason 2y agothe model is too small to memorise that much, and papers usually check things like "closest sample from training set" so we know it's not memorising (unless it's an image that's hundreds of times in the training set like the Mona Lisa)
- puttycat 2y agoHow do you quantify "by that much"? It can memorize at least some of the images for sure. Potentially these are the images OP was able to recover.
- WithinReason 2y agoWhy would it memorize some but not others if each contributes the same amount of information? E.g. the LAION dataset is about 240TB, while FLUX is about 10GB, so it could only memorize 0.00833333% of each image, not enough to reproduce it.
- DougN7 2y agoI had the exact same question.
- joegibbs 2y agoWow that's amazing, you really can't tell. What is it that gives most AI images that not-quite-real look? Is it due to airbrushed images in the training data, including cartoons, 3D renders and illustrations or all of that?
- Hasnep 2y agoThe image linked in that comment has a background that's half car and half building, but yeah apart from that minor detail it's pretty hard to tell.
- pepa65 2y agoI was going to say that. Pretty crazy! The seatbelt looks more like a robotic arm. Yet at the back there is a curve and a back window. It's really a hybrid car-building. Wondering what the black bar at the top is about...
- mkl 2y agoMost of these look very good! Signs, text, clocks, number plates, and girders or spindly branches tend to go wrong. There are a few mutant hands [0]. Mountain Sheep in Snow [1] looks fake to me - the sheep are more like dogs fading back to rocks, and have different scales. Underground Time Display [2] looks obviously fake as the clock's colon is in the wrong place and the sign has fake writing. [0] https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/ozjUKlysLqlFzajSdOdn1.jpg https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/ozjUKlysLql..., https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/XcNnWM6L54QTKp3FcH3VZ.jpg https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/XcNnWM6L54Q... [1] https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/It9DhumKX7w_2XGuPOtEY.jpg https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/It9DhumKX7w... [2] https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/GygiLirdHvqZRDmvfnMw3.jpg https://d1l4k1vcf8ijbs.cloudfront.net/fakeimages/GygiLirdHvq...