[c0decracker]

Interesting. I got access couple weeks ago (was on waitlist since the initial announcement) and frankly as much as really want to be excited and like it, DALL-E ended up being a bit underwhelming. IMHO - often results that produced are of low quality (distorted images, or quite wacky representation of the query). Some styles of imagery are certainly a better fit for being generated by DALL-E, but as far as commercial usage I think it needs a few iterations and probably even larger underlying model.

[arecurrence]

I suspect you simply need to use it more with a lot more variation in your prompts. In particular, it takes style direction and some other modifiers to really get rolling. Run at least a few hundred prompts with this in mind. Most will be awful output... but many will be absolute gems.

It has, honestly, completely blown me away beyond my wildest imagination of where this technology would be at today.

[simonw]

This book has some very good, actionable advice on crafting prompts that get better results out of DALL-E: https://dallery.gallery/the-dalle-2-prompt-book/

[andybak]

I also got access a couple of weeks ago and I can't fathom how anyone could be underwhelmed by it.

What were you expecting?

[c0decracker]

Fundamentally I have two categories of issues I see with DALL-E, but please don't get me wrong -- I think this is a great demonstration of what is possible with huge models and I think OpenAI work in general is fantastic. I will most certainly continue using both DALL-E and OpenAI's GPT3. (1) Between what DALL-E can do today and commercial utility is a rift in my opinion. I readily admit that I am have not done hundreds of queries (thank you folks for pointing that out, I'll practice more!) but that means that there is a learning curve, isn't it? I can't just go to DALL-E, mess with it for 5-10 minutes and get my next ad or book cover or illustration for my next project done? (2) I think DALL-E has issues with faces and human form in general. Images it produces are often quite repulsive and take the uncanny valley to the next level. I absolutely surprise myself when I noticed thinking that images with humans DALL-E produced lack of... soul? Cats and dogs on the other hand it handles much better. I done tests with other entities --- say cars or machinery -- and it generally performs so so with them too, often creating disproportionate representations of them or misplacing chunks. If you're querying for multiple objects on a scene it quite often melds them together. This is more pronounced in photorealistic renderings. When I query for painting-style it works mostly better. That said every now and then it does produce a great image, but with this way of arriving at it, how fast I'll have to replenish those credits?.. :)

All in all though I think I am underwhelmed mostly because my initial expectations were off, I am still a fan of DALL-E specifically and GPT3 in general. Now when is GPT4 coming out? :)

[harpersealtako]

Dalle seems to only have a few "styles" of drawing that it is actually "good" at. It is particularly strong at these styles but disappointingly underwhelming at anything else, and will actively fight you and morph your prompt into one of these styles even when given an inpainting example of exactly what you want.

It's great at photorealistic images like this: https://labs.openai.com/s/0MFuSC1AsZcwaafD3r0nuJTT, but it's intentionally lobotomized to be bad at faces, and often has an uncanny valley feel in general, like this: https://labs.openai.com/s/t1iBu9G6vRqkx5KLBGnIQDrp (never mind that it's also lobotomized to be unable to recognize characters in general). It's basically as close to perfect as an AI can be at generating dogs and cats though, but anything else will be "off" in some meaningful ways.

It has a particular sort of blurry, amateur oil painting digital art style it often tries to use for any colorful drawings, like this: https://labs.openai.com/s/EYsKUFR5GvooTSP5VjDuvii2 or this: https://labs.openai.com/s/xBAJm1J8hjidvnhjEosesMZL . You can see the exact problem in the second one with inpainting: it utterly fails at the "clean" digital art style, or drawing anything with any level of fine detail, or matching any sort of vector art or line art (e.g. anime/manga style) without loads of ugly, distracting visual artifacts. Even Craiyon and DALLE-mini outperform it on this. I've tried over 100 prompts to get stuff like that to generate and have not had a single prompt that is able to generate anything even remotely good in that style yet. It seems almost like it has a "resolution" of detail for non-photographic images, and any detail below a certain resolution just becomes a blobby, grainy brush stroke, e.g. this one: https://labs.openai.com/s/jtvRjiIZRsAU1ukofUvHiFhX , the "fairies" become vague colored blobs here. It can generate some pretty ok art in very specific styles, e.g. classical landscape paintings: https://labs.openai.com/s/6rY7AF7fWPb5wWiSH0rAG0Rm , but for anything other than this generic style it disappoints hard.

The other style it is ok at is garish corporate clip art, which is unremarkable and there's already more than enough clip art out there for the next 1000 years of our collective needs -- it is nevertheless somewhat annoying when it occasionally wastes a prompt generating that crap because you weren't specific that you wanted "good" images of the thing you were asking for.

The more I use DALLE-2 the more I just get depressed at how much wasted potential it has. It's incredibly obvious they trimmed a huge amount of quality data and sources from their databases for "safety" reasons, and this had huge effects on the actual quality of the outputs in all but the most mundane of prompts. I've got a bunch more examples of trying to get it to generate the kind of art I want (cute anime art, is that too much to ask for?) and watching it fail utterly every single time. The saddest part is when you can see it's got some incredible glimpse of inspiration or creative genius, but just doesn't have the ability to actually follow through with it.

[napier]

GPT3 has seen similar lobotomization since its initial closed beta. Current davinci outputs tend to be quite reserved and bland, whereas when I first had the fortunate opportunity to experience playing with it in mid 2020, if often felt like tapping into a friendly genius with access to unlimited pattern recognition and boundless knowledge.

[harpersealtako]

I've absolutely noticed that. I used to pay for GPT-3 access through AI Dungeon back in 2020, before it got censored and run into the ground. In the AI fiction community we call that "Summer Dragon" ("Dragon" was the name of the AI dungeon model that used 175B GPT-3), and we consider it the gold standard of creativity and knowledge that hasn't been matched yet even 2 years later. It had this brilliant quality to it where it almost seemed to be able to pick up on your unconscious expectations of what you wanted it to write, based purely on your word choice in the prompt. We've noticed that since around Fall 2020 the quality of the outputs has slowly degraded with every wave of corporate censorship and "bias reduction". Using GPT-3 playground (or story writing services like Sudowrite which use Davinci) it's plainly obvious how bad it's gotten.

OpenAI needs to open their damn eyes and realize that a brilliant AI with provocative, biased outputs is better than a lobotomized AI that can only generate advertiser-friendly content.

[visarga]

So it got worse for creative writing, but it got much better at solving few-shot tasks. You can do information extraction from various documents with it, for example.

[napier]

I mean yes, you’re right insofar as it goes. However nothing I am aware of implies technical reasons linking these two variables into a necessarily inevitable trade-off. And it’s not only creative writing that’s been hobbled; GPT3 used to be an incredibly promising academic research tool and given the right approach to prompts could uncover disparate connections between siloed fields that conventional search can only dream of.

I’m eager for OpenAi to wake up and walk back on the clumsy corporate censorship, and/or for competitors to replicate the approach and improve upon the original magic without the “bias” obsession tacked on. Real challenge though “bias” may pose in some scenarios, perhaps a better way to address this would be at the training data stage rather than clumsily gluing on an opaque approach towards poorly implemented, idealist censorship lacking in depth (and perhaps arguably, also lacking sincerity).

[whywhywhywhy]

The face thing is weird in context of them not being worried about it infringing on the copyright of art. If they're confident it's not going to infringe on art copyright, why the worry it might generate the face of a real person.

[dereg]

I felt the same way. If anything, I realized how soulless and uninteresting faceless art is. Dall-E 2 goes out of its way to make terrible faces for, im guessing, deepfake reasons?