4 ms·

I'm impressed by all of these image generators but I still don't see them working toward being able to say, "Give me an astronaut riding a horse. Ok, now the s

I'm impressed by all of these image generators but I still don't see them working toward being able to say, "Give me an astronaut riding a horse. Ok, now the same location where he arrives at a rocket. Now one where he dismounts. Now the horse runs away as the astronaut enters the rocket." You can ask for all those things but the AI still has no idea what it's doing and cannot tell you where the astronaut is, etc.

masswerk 4y ago
I'd also say, every of these images would fail a reverse test (i.e., asking a person to describe the image and what it represents.) The task is not just about generating an image that may somehow be in accordance with the prompt, but also to generate a significant image. [Edit] The equivalent to a Turing test for compositional images would be something like this: have as set of 100 images with their respective prompts, some generated by an AI, some by a human graphic designer / artist; let the test person pick the images that were generated by a computer. Mind that this would not only involve the problem of compositionality per se, but also a meaningful and/or artistic composition of the image itself. Is someone attempting to express what is given in the prompt?
- m00x 4y ago
  They actually use the reverse test to train the generator, and to score which image is most relevant to the prompt from the many images given by the generator. Dall-E does this using the OpenAI CLIP model. You can see the mini version here using this exact logic https://wandb.ai/dalle-mini/dalle-mini/reports/DALL-E-Mini-Explained--Vmlldzo4NjIxODA https://wandb.ai/dalle-mini/dalle-mini/reports/DALL-E-Mini-E...
  - masswerk 4y ago
    What I'm aiming at is about what is shown by an image, not what is in an image. Take for example the images for "A digital art picture of a robot child riding a llama with a bell on its tail through a desert", which Scott Alexander counts as a win. The first image actually shows a merry llama, a robot, which is unmistakably a robot child, riding the llama and it's clearly a desert scene. If we forget for a moment about the missing bell, it's probably the best picture. But it is also very blunt in composition. I can't imagine why anybody should have made this image. Maybe, if somebody approached a designer, like, "See, we have this wooden toy cube and need an illustration for this face of the cube. What about a cute picture of a robot child riding a llama with a bell on its tail through a desert?" – But, at closer inspection, there's something sinister going on: it's rather the llama that is leading the robot child by a rein, not the other way round. – I mean, this is meant to be a cute toy! And where is the bell? We need to talk about that contract again… The second image is undisclosed, so we can't really say anything about this. The third image is rather special. The llama seems to be robotic as well, the robot, which is – again – clearly a child, seems to be not only riding the llama, but both appear somehow integrated into a single unit, which cumulates in the robot child's face screen. There is an eerie feeling about this image. The fact, that the bell seems to be attached to the rein as some kind of link between the llama and its rider doesn't exactly help. (There's also a conic extrusion at the back of the llama, but I'd rather interpret this as part of the llama, and it's not attached to its tail.) The composition in its flat side view produces a tension focusing towards the left side of the frame, on something, which is not shown, but apparently a vital part of the story. While I might notice the mountain in the background, I'd probably forget to write home about the scene being set in the desert. But I would note that we're missing context to understand this image and what may be shown by it. The fourth image, finally, is clearly Star Wars, robot edition. However, no bell. ("A robot child riding a llama with a bell on its tail through a desert" – "Ah, you mean Star Wars!") I'm not even sure which of these images Alexander did pick as a winner. And I would describe neither image by the prompt, nor would I dare to imagine that a human had chosen these exact means to show what is described in the prompt. Having said that, thanks for the link to the DALL-E Mini paper!
- t_mann 4y ago
  I bet most humans would fail this test on images that everybody agrees are adequate portrayals. Answering a short query with an image is a highly non-injective mapping, you simply don't know what aspects of the scene were specifically asked for in the query, and which ones were filled in by the artist / AI. Eg, the queries "opening of a medieval theme park", "announcement of a witch trial", "king charles proclamation" might all be reasonably answered by similar images containing a small crowd and a speaker in a medieval-looking setting, even though they're not meant to refer to the same time periods or settings at all.
  - masswerk 4y ago
    Mind that the test is meant to include the prompt/query. E.g., take one of Alexander's winners, the robot in the cathedral: a human would probably answer to the prompt by making the cathedral part of the subject, by investing some effort in pointing out that this is indeed a cathedral, instead of just conforming to the query by showing some ambigue bits in the background, which may or may not represent parts of a cathedral. The quest of the machine is still "create an image by allotting the elements provided in the query", not "compose an image showing this and that as the subject" – and there's a significant difference. Closing this gap would require the AI to form a concept of what is given as a subject in its entirety and then constructing a plausible scene around this, by a meaningful placement of the subject, which is clearly beyond the state of art. I admit that there is a certain appeal to those images, for their distinctive dreamlike quality, as there's often a specific tension in the rather blunt composition and an apparent subject, which seems to be beyond what is actually depicted, as if this was just a casually picked specimen from a series of illustrations for a broader story line. But I'd bet that we're going to have been seeing too much of this soon, in order to be still amazed.
z3c0 4y ago
I guess you're technically correct, but the task you're describing isn't generating an image from a prompt. It would be to maintain context across distinct-but-related statements based on an internalized model of reality. That's like discounting the advent of the calculator because you still need an accountant.
Barrin92 4y ago
what shows how low level these models still are is that they don't seem to be able to draw text on a surface. It's generally just nonsense. Going higher in abstraction like asking for permanence of distinct entities or world knowledge, like having a player face the basketball hoop is several levels above that yet. I think that puts pretty severe limits on what you can do with it because in a videogame, a comic strip or basically any piece of sequential art you need to keep track of characters and environments as objects.
- sdenton4 4y ago
  Check out the model size comparison with the kangaroos here: https://parti.research.google/ https://parti.research.google/ Embedding a language model in the image generation model seemingly just requires a bigger network.
andreyk 4y ago
Technically this is possible with these same techniques if you just initialize the image with the prior one, though I am sure that does not work that well. Really you need image+text->image instead of just text->image generation. Some examples of relevant papers: "Conditioned and composed image retrieval combining and partially fine-tuning CLIP-based features", "IMAGE GENERATION WITH MULTI-MODAL PRIORS USING DENOISING DIFFUSION PROBABILISTIC MODELS". There was a more recent one I saw on Twitter I don't recall the name of. I wouldn't be surprised if these kinds of things work well by a year from now.
QuadmasterXLII 4y ago
This is just composition again: If Imagen had compositionality, it would generate the four images you want from the prompt “A four panel webcomic: first, an astronaut riding a horse. Ok, now the same location where he arrives at a rocket. Now one where he dismounts. Now the horse runs away as the astronaut enters the rocket."
- thaumasiotes 4y ago
  That is not composition in the linguistic sense. It's context. Composition will tell you that in the phrase "the same location where he arrives at a rocket", "same" modifies "location" rather than "rocket", but it won't tell you what "same" refers to.
sdenton4 4y ago
So, what you're asking for is shared context over multiple prompts, which really isn't what this generation of models is trained for. It's moving the goalposts on the mounted astronaut. However, there is progress towards what you're asking for. The recent work on textual inversion is in the right direction: https://github.com/hlky/sd-enable-textual-inversion https://github.com/hlky/sd-enable-textual-inversion It creates a representation of an entity and allows rending it in different styles and contexts. Currently it involves model fine tuning, but I expect it will become convenient as the power of the operation becomes clear. And once it's convenient, you'll be able to do the progressive queries you're asking for (and it'll be a lot easier to create narratively coherent sets of images.)
- ForHackernews 4y ago
  > which really isn't what this generation of models is trained for. Exactly. AI hypemen would have us believe that training ever-larger models on ever-larger datasets is making meaningful progress towards general intelligence, but these kind of simple tests reveal this supposed "intelligence" for what it is: fancy pattern recognition. Questions that a six year old would easily answer, these models fail at.
projektfu 4y ago
I guess what I'm saying is that I agree we're in the Clever Hans stage of AI, where we're just more explicit about stopping Hans when his tapping has reached our goal. I think the ability of these image models to synthesize new images is really amazing. It makes the computer feel like it is doing something organic, and not just applying filters and things to the images. Then, when we see the new image paired with a text that generated it, we think the system might actually know what we're talking about. But it obviously doesn't, it's just the luck of a model with billions of parameters. Whenever the model fails to produce an acceptable output, it stops being intelligent and the user is considered to be bad at their job, or to be asking for something that is unreasonable. I think it's still spot-on to say that comprehension is far away, even though you can pair outputs to inputs and have a simulacrum of comprehension.
- sdenton4 4y ago
  Take a look at the curse of dimensionality... We're at the stage of reducing a haystack from nearly infinite to a small pile of hay to search for the needle, which has required massive advances. This really isn't clever hans at all. Additionally, it's helpful to look at these systems as tools. We don't expect cars to work well without humans learning how to interact with them in a safe and reliable way. ML tools, thanks to high expectations and moving goalposts, aren't tested in the same way. But ultimately this specific line of questions - handling context over multiple queries - is something people are actively working on, and I'm confident it'll have some real solutions within a year. It's closely connected to synthesizing video, which has a huge amount of effort going in right now and some really incredible early results already. And then we can move the goal pays again and continue talking about horses...

masswerk 4y ago

I'd also say, every of these images would fail a reverse test (i.e., asking a person to describe the image and what it represents.) The task is not just about generating an image that may somehow be in accordance with the prompt, but also to generate a significant image. [Edit] The equivalent to a Turing test for compositional images would be something like this: have as set of 100 images with their respective prompts, some generated by an AI, some by a human graphic designer / artist; let the test person pick the images that were generated by a computer. Mind that this would not only involve the problem of compositionality per se, but also a meaningful and/or artistic composition of the image itself. Is someone attempting to express what is given in the prompt?

m00x 4y ago
They actually use the reverse test to train the generator, and to score which image is most relevant to the prompt from the many images given by the generator. Dall-E does this using the OpenAI CLIP model. You can see the mini version here using this exact logic https://wandb.ai/dalle-mini/dalle-mini/reports/DALL-E-Mini-Explained--Vmlldzo4NjIxODA https://wandb.ai/dalle-mini/dalle-mini/reports/DALL-E-Mini-E...
- masswerk 4y ago
  What I'm aiming at is about what is shown by an image, not what is in an image. Take for example the images for "A digital art picture of a robot child riding a llama with a bell on its tail through a desert", which Scott Alexander counts as a win. The first image actually shows a merry llama, a robot, which is unmistakably a robot child, riding the llama and it's clearly a desert scene. If we forget for a moment about the missing bell, it's probably the best picture. But it is also very blunt in composition. I can't imagine why anybody should have made this image. Maybe, if somebody approached a designer, like, "See, we have this wooden toy cube and need an illustration for this face of the cube. What about a cute picture of a robot child riding a llama with a bell on its tail through a desert?" – But, at closer inspection, there's something sinister going on: it's rather the llama that is leading the robot child by a rein, not the other way round. – I mean, this is meant to be a cute toy! And where is the bell? We need to talk about that contract again… The second image is undisclosed, so we can't really say anything about this. The third image is rather special. The llama seems to be robotic as well, the robot, which is – again – clearly a child, seems to be not only riding the llama, but both appear somehow integrated into a single unit, which cumulates in the robot child's face screen. There is an eerie feeling about this image. The fact, that the bell seems to be attached to the rein as some kind of link between the llama and its rider doesn't exactly help. (There's also a conic extrusion at the back of the llama, but I'd rather interpret this as part of the llama, and it's not attached to its tail.) The composition in its flat side view produces a tension focusing towards the left side of the frame, on something, which is not shown, but apparently a vital part of the story. While I might notice the mountain in the background, I'd probably forget to write home about the scene being set in the desert. But I would note that we're missing context to understand this image and what may be shown by it. The fourth image, finally, is clearly Star Wars, robot edition. However, no bell. ("A robot child riding a llama with a bell on its tail through a desert" – "Ah, you mean Star Wars!") I'm not even sure which of these images Alexander did pick as a winner. And I would describe neither image by the prompt, nor would I dare to imagine that a human had chosen these exact means to show what is described in the prompt. Having said that, thanks for the link to the DALL-E Mini paper!
t_mann 4y ago
I bet most humans would fail this test on images that everybody agrees are adequate portrayals. Answering a short query with an image is a highly non-injective mapping, you simply don't know what aspects of the scene were specifically asked for in the query, and which ones were filled in by the artist / AI. Eg, the queries "opening of a medieval theme park", "announcement of a witch trial", "king charles proclamation" might all be reasonably answered by similar images containing a small crowd and a speaker in a medieval-looking setting, even though they're not meant to refer to the same time periods or settings at all.
- masswerk 4y ago
  Mind that the test is meant to include the prompt/query. E.g., take one of Alexander's winners, the robot in the cathedral: a human would probably answer to the prompt by making the cathedral part of the subject, by investing some effort in pointing out that this is indeed a cathedral, instead of just conforming to the query by showing some ambigue bits in the background, which may or may not represent parts of a cathedral. The quest of the machine is still "create an image by allotting the elements provided in the query", not "compose an image showing this and that as the subject" – and there's a significant difference. Closing this gap would require the AI to form a concept of what is given as a subject in its entirety and then constructing a plausible scene around this, by a meaningful placement of the subject, which is clearly beyond the state of art. I admit that there is a certain appeal to those images, for their distinctive dreamlike quality, as there's often a specific tension in the rather blunt composition and an apparent subject, which seems to be beyond what is actually depicted, as if this was just a casually picked specimen from a series of illustrations for a broader story line. But I'd bet that we're going to have been seeing too much of this soon, in order to be still amazed.

z3c0 4y ago

I guess you're technically correct, but the task you're describing isn't generating an image from a prompt. It would be to maintain context across distinct-but-related statements based on an internalized model of reality. That's like discounting the advent of the calculator because you still need an accountant.

Barrin92 4y ago

what shows how low level these models still are is that they don't seem to be able to draw text on a surface. It's generally just nonsense. Going higher in abstraction like asking for permanence of distinct entities or world knowledge, like having a player face the basketball hoop is several levels above that yet. I think that puts pretty severe limits on what you can do with it because in a videogame, a comic strip or basically any piece of sequential art you need to keep track of characters and environments as objects.

sdenton4 4y ago
Check out the model size comparison with the kangaroos here: https://parti.research.google/ https://parti.research.google/ Embedding a language model in the image generation model seemingly just requires a bigger network.

andreyk 4y ago

Technically this is possible with these same techniques if you just initialize the image with the prior one, though I am sure that does not work that well. Really you need image+text->image instead of just text->image generation. Some examples of relevant papers: "Conditioned and composed image retrieval combining and partially fine-tuning CLIP-based features", "IMAGE GENERATION WITH MULTI-MODAL PRIORS USING DENOISING DIFFUSION PROBABILISTIC MODELS". There was a more recent one I saw on Twitter I don't recall the name of. I wouldn't be surprised if these kinds of things work well by a year from now.

QuadmasterXLII 4y ago

This is just composition again: If Imagen had compositionality, it would generate the four images you want from the prompt “A four panel webcomic: first, an astronaut riding a horse. Ok, now the same location where he arrives at a rocket. Now one where he dismounts. Now the horse runs away as the astronaut enters the rocket."

thaumasiotes 4y ago
That is not composition in the linguistic sense. It's context. Composition will tell you that in the phrase "the same location where he arrives at a rocket", "same" modifies "location" rather than "rocket", but it won't tell you what "same" refers to.

sdenton4 4y ago

So, what you're asking for is shared context over multiple prompts, which really isn't what this generation of models is trained for. It's moving the goalposts on the mounted astronaut. However, there is progress towards what you're asking for. The recent work on textual inversion is in the right direction: https://github.com/hlky/sd-enable-textual-inversion https://github.com/hlky/sd-enable-textual-inversion It creates a representation of an entity and allows rending it in different styles and contexts. Currently it involves model fine tuning, but I expect it will become convenient as the power of the operation becomes clear. And once it's convenient, you'll be able to do the progressive queries you're asking for (and it'll be a lot easier to create narratively coherent sets of images.)

ForHackernews 4y ago
> which really isn't what this generation of models is trained for. Exactly. AI hypemen would have us believe that training ever-larger models on ever-larger datasets is making meaningful progress towards general intelligence, but these kind of simple tests reveal this supposed "intelligence" for what it is: fancy pattern recognition. Questions that a six year old would easily answer, these models fail at.

projektfu 4y ago

I guess what I'm saying is that I agree we're in the Clever Hans stage of AI, where we're just more explicit about stopping Hans when his tapping has reached our goal. I think the ability of these image models to synthesize new images is really amazing. It makes the computer feel like it is doing something organic, and not just applying filters and things to the images. Then, when we see the new image paired with a text that generated it, we think the system might actually know what we're talking about. But it obviously doesn't, it's just the luck of a model with billions of parameters. Whenever the model fails to produce an acceptable output, it stops being intelligent and the user is considered to be bad at their job, or to be asking for something that is unreasonable. I think it's still spot-on to say that comprehension is far away, even though you can pair outputs to inputs and have a simulacrum of comprehension.

sdenton4 4y ago
Take a look at the curse of dimensionality... We're at the stage of reducing a haystack from nearly infinite to a small pile of hay to search for the needle, which has required massive advances. This really isn't clever hans at all. Additionally, it's helpful to look at these systems as tools. We don't expect cars to work well without humans learning how to interact with them in a safe and reliable way. ML tools, thanks to high expectations and moving goalposts, aren't tested in the same way. But ultimately this specific line of questions - handling context over multiple queries - is something people are actively working on, and I'm confident it'll have some real solutions within a year. It's closely connected to synthesizing video, which has a huge amount of effort going in right now and some really incredible early results already. And then we can move the goal pays again and continue talking about horses...