15 ms·
Show HN: This Food Does Not Exist
- jdthedisciple 4y agoLooks impressive but I can't escape the notion that surely some of the generated images will be very close to the some of the training images? How am I to assess how original the generated results really are?
- danuker 4y agoImage search, I guess. No results, it's original enough.
- kick_in_thedoor 4y ago
- hwers 4y agoYeah stylegan is rarely this good at super diverse data like this
- gus_massa 4y agoAre you using the same model for cookies and cheesecakes? Do you get sometimes a cookiecake?
- MasterScrat 4y agoWe currently train each model independently, ie we first gather a cookie dataset, train a cookie model then restart from scratch for the next one. That's actually something we're investigating: can we train a single class-conditional model for multiple types of food? Or, can we finetune cheesecakes from cookies?
- TuringNYC 4y ago>> ie we first gather a cookie dataset, Is there a chance your dataset provider makes a claim that they have derived data rights over your model generated images? Would you have sufficient confidence, say, to sell your images on a stock image site?
- zorgmonkey 4y agoIt is still somewhat unclear, but it seems that images generated by a machine learning model are not copyrightable (to quote the US Copyright Office, generated images "lack the human authorship necessary to support a copyright claim"). Whether the model itself is copyrightable is less clear to me, but [0] seems to suggest that it be. All of this depends on the country, but much of the world tends to eventually mimic US copyright law. [0] https://law.stackexchange.com/questions/19981/who-can-claim-copyrights-on-machine-learning-models https://law.stackexchange.com/questions/19981/who-can-claim-...
- dylan604 4y agoWell, now I want a cookiecake.
- MasterScrat 4y agoWe have trained four StyleGAN2 image generation models and are releasing checkpoints and training code. We are exploring how to improve/scale up StyleGAN training, particularly when leveraging TPUs. While everyone is excited about DALL·E/diffusion models, training those is currently out of reach for most practitioners. Craiyon (formerly DALL·E mega) has been training for months on a huge TPU 256 machine. In comparison our models were each trained in less than 10h on a machine 32x smaller. StyleGAN models also still offer unrivaled photorealism when trained on narrow domains (eg thispersondoesnotexist.com), even though diffusion models are catching up due to massive cash investments in that direction.
- goldemerald 4y agoI don't suppose you have a way of converting these models into a pytorch usable version, do you?
- forgingahead 4y agoNice work - can you share how large your training datasets are in terms of number of images? And did you train your models from scratch or fine-tune them from an existing model?
- andrewmcwatters 4y agoDarn! I was hoping for other-worldly foods that don't actually exist being generated from real food attributes. I suppose I should have known better.
- spacemanmatt 4y agoSomewhere in that data set is found an Eigencookie. I want the recipe.
- WalterSear 4y agoNo hot dogs? Nice work.
- dylan604 4y agoWhen will we see this as a contestant on Is It Cake?
- croes 4y agoYou are killing Instagram influencers
- Spivak 4y agoYou mean supplying. Imagine running a food IG that didn’t even need to make the food.
- scifibestfi 4y agoSeriously, won't this combined with GPT-3 flood the influencer market?
- golergka 4y agoAnd here I was, hoping for new, never seen before dishes.
- forgotusername6 4y agoI wonder how close the nearest match from the training data is. Was there a cheesecake that looked almost like these generated images?
- layer8 4y agoMaybe the ML model effectively implements a lossy image database with minor randomization. :)
- fxtentacle 4y agoSince GANs are effectively one class of denoising auto-encoders, your summary is spot-on. This type of ML model learns to effectively compress and decompress natural images by representing it as a hierarchy of convolutional features = shape templates.
- ad404b8a372f2b9 4y agoI don't think it's accurate at all to characterize GANs as denoising auto-encoders, they're not even superficially similar, unless you're talking about a very specific architecture of autoencoder-based GANs like AAEs.
- fxtentacle 4y agoThe part of a GAN that people use to synthesize images has to be a denoising decoder. You put in 1x1px high-dimensional noise as the embedding and it'll gradually upscale and turn that noise into the 3-dimensional output image. The part of a GAN that people use for control is based on an encoder. You put in your 3-dimensional source image and it gets converted to that 1x1px high-dimensional noise in such a way that putting the noise into the encoder will produce your image again. So the network architecture/structure of a GAN is the exact same structure as you'd use for a denoising autoencoder. You might train that GAN with a different intent, but it's still a stack of upscalers and convolutions that perfectly matches the architecture inside the decoder part of a denoising auto-encoder. BTW sometimes people also explicitly call this out. The VQ part in VQGAN stands for VQVAE, the "Vector-quantized Variational AutoEncoder". And VQGAN+CLIP is what the first open source DALL-E clones were based on.
- waynesonfire 4y agoand what was the licensing for the training data that you used?
- ComputerCat 4y agoEverything looks delish!
- Sebbecking 4y agoHow big was your training dataset?
- n4bz0r 4y agoThe food looks great! I suppose these models could use some extra training with dishes, though. The plates and glasses look wobbly, which is an instant giveaway. Otherwise, I can see this being used by food posters! Maybe not as a primary source, but as a "filler" — for sure.
- hammycheesy 4y agoI tried to use the linked Colab notebook to generate my own, and it appears to have been successful, but I don't see any way to view the generated images via the notebook interface. I'm not familiar with the notebook tool - have I missed something?
- sireat 4y agoIf the result is standard numpy 3d matrix then something like Pillow should be able to display the images. Something like from matplotlib import pyplot as plt plt.imshow(matrix) plt.show()
- lancebeet 4y agoI ran it locally and it generated images as PNGs in the "generated_images" directory (named 0.png, 42.png etc. after the seeds provided to the script). If it works and does the same in the notebook, you should be able to click the folder icon in the menu on the left to open the file browser, expand the "generated_images" directory, then click the ellipses next to each file to select "Download".
- munificent 4y agoI love cheesecake with strawraspcherries on top.
- derbOac 4y agoThe Cake is a Lie meme was never so relevant.
- mike_hock 4y agoAnd the Science gets done and you make a neat gun.
- beej71 4y agoWe're looking at the complete collapse of the stock photography market.
- Kye 4y agoShrinking microstock rates already killed it.
- echelon 4y agoIt's way more than that. Anyone can be an artist, musician, photographer, writer. It's going to result in more content being created, which will change the economies of content. Rate, scale, and volume of production will increase by orders of magnitude. Disney thinks IP is a war chest. That's an old way of thinking. Star Wars won't be special to the new kids growing up that can generate "Space Samurai" and "Galaxy Brouhaha" in an afternoon. We're going to hit a Cambrian explosion of content.
- smaudet 4y ago"It's going to result in more content being created" Is it, though? This model took over a month, on extremely fit hardware, to even create. Lets say for a second, in some hypothetical future, that anyone can access/use/update these models (by anyone, I mean someone with both low amount of resources as well as little to no programming skill), why are they creating content? "Rate, scale, and volume of production will increase by orders of magnitude." If by production you mean "paid creation", I'm not so sure about that. In this world where everyone creates content from thin air 1) there is little to no monetary value to the content anymore (as monetary value inversely correlates with scarcity) 2) So there is less incentive to create anything, because there is no monetary value to doing so. In fact, by definition we can pretty much prove that not much of anything will happen in this regard, because content is already limited by budget - the budget has not gone up, and the return has only gotten worse (in this hypothetical scenario). What I think is more likely to happen - a few, "blessed" individuals will have out-sized content creation capabilities, without much need to innovate. The rest of us will have almost no incentive to create anything as a result. Disney will use these systems, and they will use them to churn out more garbage, faster, on average, most kids will not be generating any movies in an afternoon.
- notamy 4y agoI was really hoping that this would be never-before-seen, AI-generated recipes or something similar ):
- sovok 4y agoThat would be the OpenAI Recipe creator (eat at your own risk) https://beta.openai.com/examples/default-recipe-generator https://beta.openai.com/examples/default-recipe-generator
- JadoJodo 4y agoOP: Forgive me if this is out of place. Also, please know that my question is genuine, not at all a reflection on the author/their project, and most certainly born out of my own ignorance: Why are these kinds of things impressive? I think part of my issue is that I don't really "get" these ML projects ("This X does not exist" or perhaps ML in general). My understanding is that, in layman's terms, workers are shown many, many examples of X and then are asked to "draw"/create X, which they then do. The corollary I can think of is if I were to draw over and over for a billion, billion years and each time a drawing "failed" to capture the essence of a prompt (as deemed by some outside entity), both my drawing, and my memory of it was erased. At the end of that time, my skill in drawing X would be amazing. _If_ that understanding is correct, it would seem unimpressive? It's not as though I can pass a prompt of "cookie" to an untrained generator and, it pops out a drawing of one. And likewise, any cookie "drawing" generated by a trained model is simply an amalgam of every example cookie. What am I missing?
- bee_rider 4y agoFor the longest time it was assumed that creativity was an almost magically human trait. The fact that somebody can, with a straight face, say "I don't get why it is impressive, I could draw these images too" is actually indicative of the wild change that has occurred over these last couple years. I guess it is true that more than a couple demos like this have been shown, so some of the awe might have worn off, but it is still pretty shocking to lots of us that you can describe the general idea of something to a computer and it can figure out and produce "what you mean," fuzzy as that is.
- JadoJodo 4y ago> The fact that somebody can, with a straight face, say ... To be clear, I'm not trying to devalue this at all; In fact, as I noted above, I am certain I'm missing something and that was what my comment was aimed at. In any case, thank you for taking the time to reply (seriously).
- bee_rider 4y ago
- gffrd 4y agoI like the thought that, years from now, we're all drinking eating weirdly-presented food / drinking weird cocktails because AI synthesized the images of drinks around the web and decided `cocktails always include fruit` and `all food must be piled high on plate`
- wyldfire 4y agoAre there any analysis techniques that can easily distinguish between these and real photographs? Do simple things like edge detections or histograms reveal any anomalies?
- daveguy 4y agoNeural networks can be trained to identify the difference, but I don't know how specific that is to the generating model. In fact, the GAN technique, at a high level is two networks -- one trying to distinguish the difference and one trying to create images that cannot be distinguished. That is the "adversarial" aspect. It is an interesting question that there may be some simple pre-processing techniques (edge detection, Fourier transform, etc) that more easily distinguish the image as a fake. Something like a shortcut from training a network to make the distinction.
- fxtentacle 4y agoI'm honestly surprised that they trained a StyleGAN. Recently, the Imagen architecture has been show to be both easier in structure, easier to train, and even faster to produce good results. Combined with the "Elucidating" paper by NVIDIA's Tero Karras you can train a 256px Imagen* to tolerable quality within an hour on a RTX 3090. Here's a PyTorch implementation by the LAION people: https://github.com/lucidrains/imagen-pytorch https://github.com/lucidrains/imagen-pytorch And here's 2 images I sampled after training it for some hours, like 2 hours base model + 4 hours upscaler: https://imgur.com/a/46EZsJo https://imgur.com/a/46EZsJo * = Only the unconditional Imagen variant, meaning what they show off here. The variant with a T5 text embedding takes longer to train.
- gwern 4y agoOr, since they are comparing to Craiyon, why not just finetune Craiyon itself? Craiyon already exists, just take it off the shelf, you don't need to retrain it from scratch, so the cost to train it from scratch on everything (which is indeed quite large) is not relevant to someone who just wants to generate great food photos.
- MasterScrat 4y agoWe haven't experimented much with Imagen, but our initial conclusions were that: - It's hard to train to a photorealistic quality (we'd be happy to be proven wrong!) - There is no strong pretrained model available yet Checking the LAION Discord, the situation doesn't seem to have considerably evolved.
- xg15 4y agoAt least with DALL-E you can be sure the food has a name. For a moment I was worried this would produce vaguely food-like images where on closer look you realise you have no idea what you're looking at - like a lot of other "this X does not exist" projects seem to do. Also a bit of cultural bias in the training is shown I think. The "pile of cookies" prompt seems to mostly generate American cookies, while e.g. a German user might be disappointed they didn't get this: https://groceryeshop.us/image/cache/data/new_image_2019/ABSB0005XOJBS_0-600x600.jpg https://groceryeshop.us/image/cache/data/new_image_2019/ABSB... :)
- fxtentacle 4y agoI thought DALL-E uses a sentence-piece encoder for the text that goes into CLIP, which would suggest that you can recombine the syllables from existing words and it'll "understand" that. So both "banana chocolate cookies" and "banacoochoconakieslade" should work.
- dalmo3 4y agoI don't think a German user writing "pile of cookies", in English, would be disappointed with "English" results. Is that any different than what you get on, say, Google? Try prompting craiyon for "Ein Stapel Kekse"* :) * Google-translated
- uptown 4y agoLike some alien sushi? https://twitter.com/khirasaki/status/1543111054460272640 https://twitter.com/khirasaki/status/1543111054460272640
- YeGoblynQueenne 4y agoNice, very alien, somewhat Giger, but not very sushi.
- herpderperator 4y agoIs there a way to trigger a fresh image on demand? That's kind of what I expect when I see a does-not-exist site.
- CSMastermind 4y agoThere's a link on the page: https://colab.research.google.com/github/nyx-ai/stylegan2-flax-tpu/blob/master/notebook/image_generation.ipynb https://colab.research.google.com/github/nyx-ai/stylegan2-fl...
- georgeburdell 4y agoThis is the most disturbing “does not exist” yet. A food blog could write itself
- hbn 4y agoThey already pretty much are. Top recipe hits on Google seem to always be from like "Southern Mama Cooking Tips" or something generic like that, and you have to scroll past 8 paragraphs of context for why this person is writing a recipe and why they like it so much, totally not to hit all the SEO sweet spots, and the full life story of this "Southern Mama" that's totally not a guy in India or a robot scraping together blurbs of text from other website.
- WordAWarning 4y agoIt's not entirely SEO, that's part of it, but it's also copyright. You can't copyright a recipe. It's information that can be freely shared. Anybody can steal it and set up their own site. You can copyright a recipe that comes with a life story. Copyright laws are weird and confusing, but it's really difficult to come up with a better solution.
- tmountain 4y agoAggregate "does not exist" website for anyone who's interested. https://thisxdoesnotexist.com/ https://thisxdoesnotexist.com/
- rkagerer 4y agoMy partner is very impressionable when she see's food in a TV show. Immediately has a craving for it. This thing is like, limitless porn for her gluttony.
- twic 4y agoThis computer has pretty poor taste in cocktails.
- Animats 4y agoComing soon to the restaurant site generator of some large delivery service. ("Picture is only for illustration purposes")
- abidlabs 4y agoYou can try out the model with this interactive Gradio demo: https://huggingface.co/spaces/nyx-ai/stylegan2-flax-tpu https://huggingface.co/spaces/nyx-ai/stylegan2-flax-tpu
- immmmmm 4y agoshould call that DeepCakes