[HN Gopher] Run your own DALL-E-like image generator
___________________________________________________________________
Run your own DALL-E-like image generator
Author : pwillia7
Score : 175 points
Date : 2022-08-08 12:24 UTC (10 hours ago)
(HTM) web link (reticulated.net)
(TXT) w3m dump (reticulated.net)
| Workaccount2 wrote:
| With the explosion of image gen AI out there recently, I feel
| confident that within a year anyone will have access to free or
| nearly free totally unchained image generation on par with or
| better than current DALLE-2.
| EZ-Cheeze wrote:
| Stable Diffusion is free and much better than DALL-E at faces.
| Not open to everyone yet but getting there.
| kuboble wrote:
| It might be better at faces but I've access to stable
| diffusion and on my prompts it performed significantly worse
| than dalle.
| yreg wrote:
| Running it is one thing, but where are we going to get the
| model?
|
| Who is going to give away their trained model (that presumably
| cost millions) to the public? And getting the blame for any
| nefarious usage that will eventually follow?
| GaggiX wrote:
| StabilityAI with Stable Diffusion
| at_a_remove wrote:
| I can only dream. While at first I was excited by the mini and
| the mega versions of DALL-E 2 people were hosting, even if they
| were blurry and often a little dumb, by the time I got my DALL-E
| 2 invite, I had _ideas_ ... most of which were swiftly rejected.
| Not only that, it suggested that I would lose my asking
| privileges if I kept it up with the weird prompts.
|
| Sigh, when will I have a nightmare generator of my own?
| donkeyd wrote:
| > it suggested that I would lose my asking privileges if I kept
| it up with the weird prompts
|
| This makes me hella curious about the type of stuff you were
| requesting.
| Miraste wrote:
| I expanded on this in another comment, but their content
| policy is _incredibly_ restrictive:
|
| https://labs.openai.com/policies/content-policy
|
| There are a number of other rules that can only be discovered
| by combing their Discord or running into them headfirst.
| at_a_remove wrote:
| My first request was a Captain Beefheart lyric ("making love
| to a vampire with a monkey on my knee") because I wanted to
| see how it would take something concrete "monkey on my knee"
| and integrate it with not only a metaphor, but one that had
| changed over time ("making love" used to refer making _out_ ,
| a.a.a. prolonged kissing, rather than sex). Additionally, I
| wanted to see if it had hoovered up a rendition of that exact
| lyric off of Exploding Dog (an artist makes odd drawings from
| submitted phrases and prompts) some twenty-two years prior.
|
| Of all of the celebrities, it seemed reticent to draw Fred
| Rogers as, well, Fred Rogers, despite me being able to make
| some fairly appalling melds of others, like Rod Serling and
| Marilyn Monroe.
|
| Additionally, someone has put in an overly-touchy anti-vore
| filter.
| margoguryan wrote:
| I've tried to do a bunch of prompts involving, like,
| pirates and cannons - rapiers and daggers. 16th century
| Treasure Island aesthetic. No bueno.
|
| I also tried uploading a film still, and it said I may not
| use realistic faces. So I used another AI tool to remove
| the face. It still wouldn't budge.
| tiborsaas wrote:
| It's really easy to find a false positive prompt so you have
| to be really careful.
|
| It puzzled me for a while, but then it made sense:
|
| "A hairy ball with lights at the end is manipulated by a
| human"
|
| Not a great prompt, but it seemed like a good start. I
| imagined fiber optic strings attached to a ball and a human
| doing some stuff around it like a wizard or something. But
| then I realized that "hairy ball" might lead to something
| completely different :)
|
| Okay, no problem, I changed it to "tumbleweed" but that
| triggered the drug filter so I gave up and tried something
| else :)
| at_a_remove wrote:
| There's something hilarious about having all of this "such
| wow, much AI" machine-learning safeguarded by primitive
| wordfilters that get Scunthorped like chumps.
| GaggiX wrote:
| Stable Diffusion will be available soon and they will release
| it open source
| illuminati1911 wrote:
| "Run your own text to image prompts with CUDA, a bunch of disk
| space, and an insane amount of memory."
|
| The "insane amount of memory" turned out to be 32 GB or more. I
| thought we are talking about 512 GB or something like that.
| pwillia7 wrote:
| including the GPU memory I say it's insane :) -- Also, you need
| more than 32GB of RAM to follow the instructions verbatim in
| the repository. People were saying the scann indexing fails
| even with 64GB of RAM.
| Tepix wrote:
| Many AM4 boards that cost 70EUR or so support 128GB RAM.
|
| 128GB = 4x 32GB DDR4 UDIMM costs less than 360EUR.
|
| That's a lot less than the GPU.
| throwaway743 wrote:
| iirc any CUDA requirement requires an Nvidia GPU
| matsemann wrote:
| What I hate about the current ML space is that half the time is
| spent trying to get dependencies installed and set up correctly
| to match whatever project you're trying to run expects.
| Everything is just so brittle. Following tutorials like these
| never work out of the box.
| travbrack wrote:
| I recently went through this and conda handled the python deps
| pretty well, but the nvidia cuda stuff was a massive pain to
| get the right things installed. I think it's safe to blame
| nvidia.
| tbalsam wrote:
| > conda
| ralusek wrote:
| Node has such an easier ecosystem than python IMO. I really
| hope the move of ML from python to JS continues.
| QuadmasterXLII wrote:
| What's node's GPU acceleration story?
| ralusek wrote:
| https://www.tensorflow.org/js/guide/nodejs
|
| Tensorflow JS running on GPU/CUDA in Node.
| curiousgal wrote:
| Does npm still create 10000 directories?
| speedgoose wrote:
| No. Only 1000 these days.
| ralusek wrote:
| npm downloads the dependencies that you install.
| Sohcahtoa82 wrote:
| Easy? Maybe.
|
| Sane? Absolutely not.
|
| Though I suppose it's not really Node's fault that developers
| are importing modules like "leftpad".
|
| What's really fucking insane though is that there are modules
| like "trim-newlines" [0] that exist merely to trim \r and \n
| from the beginning and end of a string...and that this is
| such a hard task to get right that it's in version
| 4.0.2...and that a previous version had a security
| vulnerability [1].
|
| [0] https://www.npmjs.com/package/trim-newlines
|
| [1] https://www.cve.org/CVERecord?id=CVE-2021-33623
| zamadatix wrote:
| As much as I hate to say it as I hate how these kinds of
| packages attach themselves to large projects to inflate
| download numbers for a resume I think most of us here would
| probably have created the same CVE doing it naively. It was
| a regex DoS due to exponential runtime not something obtuse
| like extra bloat being poorly made.
| whywhywhywhy wrote:
| I feel like this is getting way worse the past 3 years and I
| also feel that's intentional.
| tbalsam wrote:
| I'm not so sure about this...I guess maybe there's some project
| specific stuff, but the good projects are semi stable and
| oftentimes it's just PyTorch code that you can pull and use
| elsewhere if needed.
|
| Other people's infrastructure always seems bad to me, as I'm
| sure mine does to others. The cycle and circle of life! :D
| jimbob45 wrote:
| _What I hate about the current ML space..._
|
| You're implying it isn't like that for every
| language/framework. I'd love to know what you're working in
| where the dependencies aren't always a problem. Even Carmack
| has complained about this.
| matsemann wrote:
| Never had an issue running a random java project, for
| instance. Either because people package all dependencies. Or
| because maven, while not popular, mostly just works.
|
| Of course, ML is a bit different in that it often needs more
| drivers/gpu setup. But there are loads of ecosystems that
| don't assume you have package X installed on your system,
| like python's does.
| dvh wrote:
| It was exactly the same with voice recognition and synthesis
| software. Those are complex apps, often born in academia with
| obscure dependencies.
| 8f2ab37a-ed6c wrote:
| What is the solution? Is there a reason why things are not
| better? Can all of the deps be loaded in a Docker image and be
| done with it?
| Kuinox wrote:
| > What is the solution?
|
| There are languages ecosystem that solved these issues and
| never have a problem with it. Look at it, and copy their
| tooling.
| dismantlethesun wrote:
| What are some of those ecosystems?
| Kuinox wrote:
| .NET with nuget works flawlessly, compared to bad package
| managers.
| wizofaus wrote:
| It's pretty good but certainly not flawless. Admittedly I
| probably don't have high standards. Cocoapods is one of
| the few I'd consider objectively "bad", though apparently
| among iOS developers it's considered one of the better
| ones (!).
| Kuinox wrote:
| That's why I said "compared to bad package managers."
| NuGet works, have flaws, sometimes doesn't works, but is
| usually reliable.
|
| On the other hands, I lost a lot of time with npm, and
| also a lot with pip.
| mattkrause wrote:
| Julia's package manager has basically never failed on me.
| dismantlethesun wrote:
| Thanks. That's something to look into then.
| adgjlsfhk1 wrote:
| Note that the Julia package manager (while very good) is
| very similar to (and largely inspired by) the rust
| package manager.
| matsemann wrote:
| One thing is python's dependency management is just insane in
| itself. When you then additionally have to have it install
| basically drivers for the gpu, compile loads of native code
| etc it becomes a hot mess. And half the ML things you want to
| try are academic experiments not really made for
| distribution, so they were made to work and that's it. So if
| you have some different computer setup, minor version
| mismatch of a dependency etc it will just break. Datasets or
| models excist on some url that only works for a month.
|
| It's a shame, I think the field is one where reproduction of
| results should be really welcome and feasible.
|
| Yeah, a docker image would be nice. Even a Dockerfile, even
| though that in itself may not guarantee reproducibility if
| you try to build it later. (And may have issues with gpu
| drivers etc). But at least it documents all assumptions about
| the setup.
| cellis wrote:
| No. This would take a very long post to explain why, but in
| short, it depends heavily on which NVIDIA Cards / CUDA
| drivers / versions of Linux you're using. Or you're running
| on AWS GPUs or TPUs and paying a lot more either in the form
| of straight up dollars or optionality.
| zionic wrote:
| What I can't stand is the incessant paternalism that seems to
| surround AI from big tech,
|
| I got access to DALL-E a few weeks back (after a waitlist...
| come on) and tried it out. They want all this personal info to
| even access it, and then they have this oppressive "content
| policy" that removes anything remotely fun.
|
| For example I tried "Trump riding a velociraptor on mars
| fighting aliens" because why not? Sounds hilarious. Turns out
| _any_ query with the word Trump is banned and I got a warning
| about "repeated violations might remove my access". It's not
| just trump, anything remotely non-corporate-friendly is heavily
| filtered. Don't you want to just make images of a cute teddy
| bear made of pizza instead?! I'm just so tired of it all. It's
| like the puritans of the 1990's won, but they're corporations
| now.
|
| Personally I think they can shove that nonsense up their ass.
| ricochet11 wrote:
| fwiw i found you can misspell words to get around the
| filters, e.g. try something like Dnoland Tmrump and see if it
| works
| kingkawn wrote:
| My account got deleted because the word katana was in my
| image
| astrange wrote:
| They unban you if you email them. "longsword" is an allowed
| word though.
| kache_ wrote:
| Don't worry, we'll get much better models at much lower costs
| (midjourney, stablediffusion) that you can run on your own
| hardware, and avoid the moral censorship that the
| paternalising non productive members of these companies want
| to subject the world to
|
| Pepe is banned. Why? It doesn't matter. Someone will cut
| around, and they'll get my compute instead. And I'll get my
| autogenerated pepes, regardless of whether some non
| productive AI "ethicist" thinks a cartoon frog meme is
| offensive.
| yreg wrote:
| Asking users to avoid creating fakes of actual people
| "removes anything remotely fun"? Okay...
| dismantlethesun wrote:
| I think they have a blanket no celebrities or real living
| people policy for their engine. I think that's fine since
| they don't want to be called out for helping produce
| defamatory content.
|
| They just want to avoid bad PR. That's sound safe.
| Miraste wrote:
| It is to avoid bad PR, but in the name of marketing they've
| produced the worst content policy I've ever seen:
|
| https://labs.openai.com/policies/content-policy
|
| Highlights include:
|
| -All images must be "G-rated"
|
| >Any prompt referencing politics, violence, sexuality, or
| even the concept of "health" is banned
|
| >This does occasionally include the very same LGBT themes
| their "hate" rule ostensibly protects, although not
| consistently, so there's no way to know whether an LGBT
| prompt will cause an account strike
|
| >In practice the list of bannable offenses is much longer
| than this, and includes everything from the concept of
| death to anything violence or politics-adjacent (e.g.
| nothing about war, conflict, or any kind of weapon)
|
| >OpenAI refuses to share this list because it's part of a
| "contextual" filter, even though it demonstrably bans words
|
| They advertise DALL-E 2 as artistic, but it can't make art
| with restrictions like this. The best it can do is
| corporate content farming, and even then there's no way I'd
| make part of my marketing pipeline depend on a service this
| fickle.
| astrange wrote:
| In my opinion, if you haven't managed to get DALLE to
| generate fetish art, it just means you don't have enough
| fetishes.
| tiborsaas wrote:
| Not sure if this is unique to ML, I had the same experience
| with C++, Rust, Linux tutorials and and a bunch of other
| technologies.
| phaedrus wrote:
| I definitely have the same experience trying to get C++ game
| dev libraries running. Trying to wrangle a bunch of open-
| source library dependencies into mutually acceptable versions
| whilst also keeping up with (intentional, or unintentional)
| operating system breaking changes is a never ending steeple
| chase.
| breakds wrote:
| Which is why I am using Nix to maintain my development for each
| machine learning research project.
| substation13 wrote:
| I'm interested in how this works with hardware specific code?
| Can Nix detect host hardware and build for that?
| divan wrote:
| That's the reason why I avoid using Python as much as possible
| (while I really like the language). Another reason is massive
| stacktraces everywhere instead of a meaninful user-facing error
| messages, but that's a smaller cultural thing than dependencies
| hell.
| liminalsunset wrote:
| Seems like a decent amount of training data does end up making it
| into the output - some of those images have a barely legible
| "shutterstock" white footer in them or something
| upupandup wrote:
| Asking for a friend, can this be used for porn she wants to know.
| speedgoose wrote:
| Yes, but you will see things you don't want to see.
| NicoleJO wrote:
| Friendly reminder: If you're considering incorporating someone
| else's work into your new work, you need to be aware that you may
| be violating the copyright to the original work.
| https://www.legalzoom.com/articles/what-are-derivative-works...
| epivosism wrote:
| Possibly useful: midjourney offers image creation via discord,
| 200 images/10$ or a subscription plan for more. It's great fun
| and possibly is an easier way to test out the tech.
|
| AI art notes in general: * this will permanently
| raise the standard of "programmer art" in prototype UIs, for
| example. * I expect large-scale AI image creation social
| games will take off very soon * utilities to listen to your
| conversations and illustrate people's points with poetic
| interpretations on screens next to you with captions. *
| this will/should be integrated with emoji / emoji kitchen-like
| systems - imagine iterating on a custom emoji-like response in
| snapchat! * trademark/copyright wars over input data may
| ruin this
| ev1 wrote:
| I'd be happy to pay but my god at least give me a rest api to
| curl by hand and not discord of all things.
| irq-1 wrote:
| Wow. Great ideas, thanks for sharing. I'd like to see
| image>text ai so it can explain what it's seeing, for driving,
| etc...
| yonixw wrote:
| Currently free trial [1] of 25 images if anyone wants to try...
|
| But I would also suggest [2] which is the one used in the
| article already deployed for free
|
| [1] https://discord.gg/midjourney
|
| [2] https://huggingface.co/spaces/multimodalart/latentdiffusion
| Voloskaya wrote:
| Alternatively, if you don't have a (good enough) GPU, you can
| also run it on Google Colab using one of the many available
| notebooks.
|
| For example, this one [1] works very well.
|
| Google Colab comes with a fair amount of GPU time for free.
|
| [1]:
| https://colab.research.google.com/drive/1TBo4saFn1BCSfgXsmRE...
| ionwake wrote:
| Hi, Sorry if this is extremely ignorant. But I was wondering the
| other day... are we 100% sure they arent feeding images from a
| search engine index into the ML data for Dalle?
|
| It just looks surprisingly like its mixing and matching the top
| returned image searches from an index.
|
| I'm only saying it looks like that - not that it is ofcourse. I
| don't want to undermine anyones works here, I was just wondering.
| knicholes wrote:
| You can read the whitepapers for Flamingo, Dall-E 2, Imagen,
| and Parti to see how diffusion networks and GANs work to create
| these images. I wrote up two huge paragraphs trying to explain
| it simply, but then I realized that I don't understand exactly
| how they work, either. Best source is the published research.
|
| Most networks have to do with large language models, text
| embeddings, and image embeddings.
| adamsmith143 wrote:
| This is not the case. If you want go ahead and try to generate
| pictures and then find the supposed source image it came from.
| Good luck.
| zimpenfish wrote:
| I've got the LD repo installed and use it alongside VQGAN-CLIP
| and ... VQGAN is pretty much always the better result (albeit
| much slower.) I -think- this is because the (tagging?) that the
| LD is using isn't as comprehensive. Like it doesn't seem to know
| what "brogue" means and generates nonsense with attempts at the
| word "brogue".
|
| (Caveat: I have updated neither repo since about November because
| it took a faff to get them working and I do not want to touch
| them again.)
| margoguryan wrote:
| CLIP is my favorite of all of them. But it's not especially
| friendly - like riding a horse bareback. I've made some
| ridiculous, very bizarre and interesting stuff with CLIP that
| just isn't possible with DALL-E.
| frankzander wrote:
| TBH ... I absolute hate this AI generated strange images. Most
| folks are celebrate this technology but I don't know what it's
| really for other than confusing brains. I always ask myself "is
| this real" ... I don't like to ask myself if something is real or
| don't like to be confused. Is anyone else out there who feels the
| same?
| ev1 wrote:
| I'm working on an indie game and I get surprisingly decent
| results if I ask for a basic object (think 300 mutations of a
| key or book) and downsample it to the 48x48 sprite sheet size.
| They look weird as hell at normal resolution but as pixel art
| with specific resize sampling methods it's a lot less weird,
| and good for rapid prototyping.
| dron57 wrote:
| That's such an awesome use case! Do you have any examples you
| can share?
| ev1 wrote:
| https://upload.vaa.red/Ct8Ui#48b916
| https://upload.vaa.red/bVBvk#c275ba
|
| The originals look like malformed mutants at 256x256 with
| nonsensical key bits and strange twisted handles.
|
| Note that this does not replace a good pixel art artist
| either - for things like walls, anything that needs to
| connect together or tile like a dungeon wall or castle, or
| a cohesive art style, this will not do. But for rapidly
| identifiable different quest item drops it's not terrible.
| cyborgx7 wrote:
| That is very cool and a clever use the technology. Any place
| where someone can follow your development efforts?
| ev1 wrote:
| Haha, I'm not even partway into picking a game engine. I
| was going to try out Unity but then the whole "bought by
| adtech" thing happened.
| highspeedbus wrote:
| I think AGI, when it inevitably arrives, will be so disruptive
| to our brains as sugar and sedentarism. We've not evolved to
| question whether what we see, hear and feel is real.
| manimino wrote:
| When computers started beating humans at chess, there was a
| similar moment of anxiety.
|
| One good quote from that era: "You should be no more concerned
| that a computer can beat you in chess than that a car can beat
| you in a race."
|
| Technology changes the experience of being human. Chess used to
| be the marker of human intelligence; that passed. Now some
| forms of creativity will too.
| frankzander wrote:
| I'm not anxious about it. It's just not pleasant to see.
| There are some painters like Dali (I assume Dall-E is not by
| accident) or Bosch who had created kinda similar "products"
| but to me they are interesting and fun to look at. The AI
| products are kinda useless creap, mashups without sense from
| dump machines and that's give me this wired feeling which
| makes me hate it.
| nonbirithm wrote:
| Yes, I agree. I believe that technology is like pollution,
| irrespective of its net utility on society: once it's been
| invented, it's practically impossible to go back to the state
| in history where it wasn't invented. Pretty soon it may begin
| to be the case that the invention of technologies that appear
| interesting on the surface have unexpected societal effects
| when they become mainstream, and it will be too late once the
| data and knowledge has spread to millions of individual hard
| drives.
|
| In the case of DALL-E and GPT-3, I believe they will undermine
| human creativity and usher in a new era where fewer people care
| about slowly crafted skills like painting that require years or
| decades of practice and patience to master if the barrier to
| just have the art in front of you in seconds becomes so low.
|
| It might get to the point that people will divide themselves on
| ideological grounds that AI art is "impure" and attack each
| other if they cannot prove its origin. I'm not saying that I'm
| one of those people that would join the pushback, given that a
| coming explosion in AI art is all but inevitable, but I'm
| describing what I think the new technology is going to cause
| the masses to believe - the population that aren't enthusiastic
| tech evangelists. When the expectations of millions of people
| are set by DALL-E and the like, I don't think the prospects
| will be universally positive.
|
| Look at the number of people on HN asking if certain comments
| were written by GPT-3; they seem to appear weekly. I don't
| think implying that you didn't actually write the comment you
| posted will be taken very well by some people outside of an
| insular circle like HN once the general public becomes fully
| aware of AI, and could very well grow into a well-known insult
| if there ever comes to be a rift in opinions around AI art.
| detritus wrote:
| I'm not with you, at all, but I did have a knot in my stomach
| the otehr day reading a quote from one of the founders of
| MidJourney, David Holz who apparently said
|
| "Within the next year or two, you'll be able to make content in
| real time: 30 frames a second, high resolution. It'll be
| expensive, but it'll be possible. Then, in 10 years, you'll be
| able to buy an Xbox with a giant AI processor, and all the
| games are dreams.".
|
| As someone who used to take quite a lot of psychedelics,
| there's something quite terrifying about the promise of this
| premise - it takes me back to the wrong sort of trips where the
| ever-unfolding strata of reality became too much bear and I'd
| end up mentally cowering beneath the unrelenting bigness of it
| all.
|
| The infinite dreamspace is unholy-big and nt somewhere I'd much
| choose to get lost.
|
| Or maybe I totally will...
|
| - ed - Notwithstanding the obvious realisation that I could
| just take the helmet off, or remove the contact lenses, or
| whatever we have in a couple of decades.
| thatcat wrote:
| Unless the neuralink implant works out to be a controller and
| visualizer...
| swamp40 wrote:
| Once you can complete a feedback loop, things will get
| _very_ strange _very_ fast.
|
| You will be able to visualize your own nightmares when you
| are awake.
| Workaccount2 wrote:
| At least for me, the intensity of a psychedelic trip was
| little about the visuals and a lot about that indescribable
| consciousness change. I don't think AI's are going to be
| collapsing people's innate sense of reality (the way that
| psychedelics does at least), perhaps leaving them questioning
| what media is and isn't real though.
| anon2020dot00 wrote:
| There is no need to worry about the threat of AI. While it is
| true that AI has the potential to cause great harm, it is also
| true that AI has the potential to do a great deal of good. As
| long as we are careful and responsible in our development and
| use of AI, there is no reason to believe that it will be
| anything other than a positive force in the world. -- generated
| by OpenAI from a prompt
| frankzander wrote:
| Yes I assumed it from the first few words that's kind of a AI
| product.
| jeroenhd wrote:
| I assume most images and videos on the internet are doctored
| and have been since the invention of photoshop.
|
| I don't feel the need to check for reality. I've seen
| screenshots from games that look more realistic than some
| pictures I've taken (Forza on max settings at the right angle
| might as well be a photograph) and this trend will only
| continue.
|
| This tool can be considered more of an automated version of
| meme communities, where dedicated members will spend an hour
| photoshopping muppets into historic events and making the
| pictures look absolutely believable. The only novelty I see is
| that the computer now does a lot (but not all) work for you.
|
| There are nice opportunities here. If you need a stock photo of
| something very specific, you'll soon be able to generate one
| with the right query and the right AI. Small companies can
| generate fancy brands without paying professional designer's
| fees, especially if all they need is a billboard and not a
| whole suite of office supplies. You can generate your own
| posters and decorations featuring interesting landscapes and
| scenes in any style you want.
|
| The current iterations of these algorithms are quite limited in
| many aspects and sometimes uncanny or even horrifying, but I
| look forward to a future where I can imagine something,
| describe it, and have it rendered into digital art just like I
| pictured, without having to spend decades on honing my skills
| as an artist.
| frankzander wrote:
| It's not about fake images trying to show reality but it's
| about images crafting something from reality and showing
| something unreal. The brain want to interpret it and fail.
| This fail drives me nuts.
| kurishutofu wrote:
| Anyone know how this compares to min-dalle?
| andybak wrote:
| You can also use Visions of Chaos on Windows which has semi-
| automatic installer for several dozen models:
| https://softology.pro/vocrev.htm
|
| (I'm a Patreon supporter but otherwise have no connection)
| simonw wrote:
| I tried doing this on my own laptop a few weeks ago using Hugging
| Face diffusers - https://github.com/huggingface/diffusers
|
| Here's the code: from diffusers import
| DiffusionPipeline ldm =
| DiffusionPipeline.from_pretrained( "CompVis/ldm-
| text2im-large-256" ) output = ldm(
| ["a painting of a raccoon reading a book"],
| num_inference_steps=50, eta=0.3,
| guidance_scale=6 )
| output["sample"][0].save("image.png")
|
| This took 5m25s to run on my laptop (no GPU configured) and
| produced a recognisable image of a Raccoon reading a book. I
| tweeted the resulting image here:
| https://twitter.com/simonw/status/1550143524179288064
| can16358p wrote:
| This looks lovely and I absolutely love where this is going.
|
| However, it would be very cool if deep learning tools weren't
| dependent on CUDA and could be run on, say, any GPU with OpenCL.
| blagie wrote:
| It would be lovely if anyone other than NVidia made GPUs
| useable for GPGPU. I started with Radeon for the open-source
| factor, but it turned out to be useless in virtually every
| respect. I bought an NVidia card.
|
| I'm now tied to CUDA, but I didn't need to be. I was starting
| from scratch.
|
| The longer before AMD can ship something which actually works,
| the more entrenched NVidia+CUDA become.
|
| At this point, I lost hope.
| atwood22 wrote:
| Would targeting Vulkan instead of CUDA help?
| naavis wrote:
| They don't really serve the same purpose. Vulkan is a low-
| level graphics API while CUDA is a much higher level
| toolkit for GPGPU.
| blueblob wrote:
| It's not that they aren't trying, it's that when you reinvent
| the wheel, you have to do more work. Microsoft is introducing
| `tensorflow-directml` to avoid this problem by implementing a
| CUDA equivalent in directX. AMD has ROCm, but it's not well
| supported because it's not integrated upstream in
| `tensorflow`.
| blagie wrote:
| Well, I bought a ROCm GPU for GPGPU:
|
| - I found out I could only use it for compute headless.
| WTF?!?!?! (https://www.phoronix.com/news/Radeon-ROCm-Non-
| GUI). If it was driving a monitor, my machine would crash
| hard. There wasn't even an error message.
|
| - A lot of other stuff didn't work and just resulted in odd
| crashes, or worse performance than CPU. I don't know why.
|
| - Within 9 months, AMD discontinued support for my card. I
| raised this as a warranty issue (suitability for advertised
| purpose), but that obviously would go nowhere without a
| lawsuit. I had a very expensive brick.
|
| - AMD support channels were non-existent. There literally
| was no way to reach anyone.
|
| I bought an NVidia card, and it's been working well ever
| since.
|
| ROCm is not well-supported because it's absolute garbage.
| You have *less* work reinventing wheels, since it's been
| invented once. You have more work to get community support
| and network effects, since you're starting out behind.
| Fundamentally, though, that can't start to happen if your
| system doesn't work at all.
|
| I agree with you they're trying, but they're trying
| incompetently.
|
| If ROCm was half the speed of CUDA, and wasn't integrating
| into the latest-greatest frameworks, but it was stable and
| working, I'd make it work. It wasn't anywhere close to
| stable and working.
| MonkeyMalarky wrote:
| More entrenched? The first time I ever touched CUDA was 2012
| in university! It's been entrenched for a whole decade!
| zepmck wrote:
| You blame CUDA because you have never played AMD + ROCm.
| Furthermore OpenCL is on dead-end street. With standard
| programming languages natively supporting CUDA, it has no sense
| to add extra layers.
___________________________________________________________________
(page generated 2022-08-08 23:01 UTC)