11 ms·
Charl-e: “Stable Diffusion on your Mac in 1 click”
- selfsimilar 4y agoWhat's the difference between this and Diffusion Bee besides a nicer website? https://github.com/divamgupta/diffusionbee-stable-diffusion-ui https://github.com/divamgupta/diffusionbee-stable-diffusion-...
- okdood64 4y agoThere's a link to a code on the `nicer website`: https://github.com/cbh123/charl-e https://github.com/cbh123/charl-e
- abc_lisper 4y agoActually, diffusion bee has code too in the parent’s link, and got a nicer website here: https://diffusionbee.com/ https://diffusionbee.com/
- mig39 4y agoIs there a reason it won't work on an intel Mac?
- zodo123 4y agoPeople have been focused on getting stable diffusion running well on M1 macs because their graphics systems have so much more horsepower than the Intel macs. The M1s also have a fast memory sharing architecture for graphics, and this needs an absolute minimum of around 8gb of vram — many Intel macs just won’t be able to handle this.
- wasyl 4y agoSD on an Intel mac with Vega graphics runs pretty well though — I think it ran at something like ~3-5 iterations/s for me, which is decent. I ran either https://github.com/magnusviri/stable-diffusion https://github.com/magnusviri/stable-diffusion or https://github.com/lstein/stable-diffusion https://github.com/lstein/stable-diffusion which have MPS support
- anigbrowl 4y agoThat's good to know as I just got a good deal on one and was wondering if the AMD GPU would be useful or if I needed to start planning for an eGPU with some NVidia silicon. Thanks!
- halefx 4y agoFrom the website: > Will this be available on Intel Macs? > Yep, I'm working on making it compatible with older Macs.
- murkt 4y agoHas Stable Diffusion been optimized already so it could run on M1 with 8 GB of RAM without swapping?
- cowmix 4y agoI highly doubt it. It Struggles on GPUs with 6GB or less.
- city17 4y agoHow fast does Stable Diffusion run on an M1 Max? I'm using an M1 Pro and I find it too slow. I'd rather use an online service that costs $0.01 per image but generates an image in a matter of seconds than wait 1 minute for a free one.
- cammikebrown 4y agoIt takes 20-30 seconds on my M1 Pro with 32GB RAM. I’m not sure I’ve seen anything faster online.
- pavlov 4y agoI clocked 18 seconds for a 512*512 image, 25 steps, on a Mac Studio with M1 Max and 32GB.
- culi 4y agoFWIW I don't I've ever gotten a satisfactory result from anything less than 50 steps
- pavlov 4y agoInteresting. I didn’t see an essential difference with higher values, so I settled on 25. Maybe I’m just impatient and my brain prefers more options even if they’re individually imperfect.
- brnaftr361 4y agoDepends on the diffuser you're using, the _a (ancestral) diffusers are aberrant in that they can yield good results with very low sample counts. I typically use 16 samples and get reasonably good results, but it's highly dependent on the prompt and settings as well.
- city17 4y agoOn Replicate it takes maybe around 10-15 seconds. My M1 Pro only has 16GB and 8 cores and takes about 1 minute, so maybe the lower specs make quite a difference.
- wasyl 4y agoIs there some comprehensive source about how to make the most of Stable Diffusion? I find the examples on websites much better than what I've been able to generate — they more closely convey the prompt and have less artifacts/clearly messed up parts
- gedy 4y agoAgreed and wondering myself, DALL-E seemed to do a better job of great looking images with brief prompts, but Stable Diffusion seems to need more specific prompts. SD is free though so would love to use it more.
- astrange 4y agoCLIP-guided Stable Diffusion, or Dalle+SD, are both doable with current open source and will have much smarter prompting at the cost of even more memory use.
- nickthegreek 4y agoI search Lexica.art for the style I want, copy the prompt associated with the work and edit it my needs.
- Nadya 4y agoWhen people call themselves "prompt engineers" it's only half in jest. Half of generating something good is guiding the program into generating something good. That means knowing the right keywords to get specific styles or effects, a little bit of luck, and sometimes generating a prompt several dozen times and then creating variations from a seed once you find a specific seed that generated something close to what you liked. It's an iterative process and many of the fantastic images you see weren't "first generations" but likely the 20th or so generation after tons of trial and error working around a specific prompt/idea. I'd recommend keeping a prompt list and finding what does/doesn't work for what you're after. Try shuffling the order of your prompt - the order of the tokens does matter! Repeat a token twice, thrice, hell make a prompt with nothing but the same token repeated 8 times. Play around with it! If you find an image that's very close to what you want - start generating variations of it. Make 20 different variations. Make variations of the variations you like best. Also the seed is very important! If you find a seed that generated a style you really liked take note of it. That seed will likely generate more things in a similar style for similar enough prompts. It's a semi-creative process and definitely takes some time investment if you want great results. Sometimes you strike gold and get lucky on your first generation - but that's rare.
- dr_dshiv 4y agoSpeed has a massive effect on how willing I am to play around and develop better prompts. I can’t wait a full minute for an image, I just can’t. What kind of computer specs would be required to generate typical SD images in less than a second?
- cstejerean 4y agoI don't know about less than 1 second but I just picked up an RTX 3090 Ti now that they're basically half off at Best Buy and it's definitely fast enough for interactively playing with prompts (single digit number of seconds). Probably overkill and could get away with something like a 3060 or so, but the 24 GB of VRAM come in handy if you want to generate larger images. I pushed it as high as 17 GB on some recent runs.
- selectodude 4y agoThe nice thing about the M1 is that the GPU and CPU share RAM so even though I have a 14" MacBook Pro, I also have a GPU with 16GB of VRAM. I pushed as high as 11GB on images and the fan didn't even turn on.
- nl 4y agoIt is slower than an NVIDIA you though. Maybe 30s per image on my M1 Max with 32gb
- midwestemo 4y agoHave a 3060 and it's fine for me, took me ~8-9 secs to produce it at default settings
- sdflhasjd 4y agoRoughly 4-5 seconds for 512x512 at 50 samples on a 3090 Ti
- skybrian 4y agoIt takes a few seconds (haven't timed it), but I suggest doing it online at dreamstudio.ai. Paying about one cent per image isn't so bad.
- gcau 4y agoA.I projects (and maybe all python projects in general) seem to always be ridiculously tedious, error-prone to get running, such that its a rare, celebratory thing when someone releases something that's easy to use like this?
- EamonnMR 4y agoMost Python projects aren't this tough. I suspect that they're using wonky libraries like Pandas, Numpy or some such that prioritize raw power over ease of installation.
- CodeSgt 4y agoI've not touched Python in a couple years, but Pandas/NumPy used to be the defacto libs for anything to do with data science, are they considered "wonky" now?
- machinekob 4y ago2be honest numpy is ez to install on all major platforms. In deep learning I'm almost never saw usage of pandas but deep learning models have problems with PyTorch as some projects just lock old PyTorch version that just dosent work on new/old python version.
- gbear605 4y agoIn my experience, they’re simultaneously important for data science and very annoying to install and manage.
- Fomite 4y agoI mean, there are at least two different companies (Enthought and Continuum) that were founded on making the major scientific python packages easier to install.
- version_five 4y agoJust to be clear, pandas and numpy are not the "wonky" libraries. They are, in my experience, basically two of the most easily installed and dependency managed libraries in python, given their ubiquity and maturity. Maybe there are machine configurations I'm not familiar with that they are not easily compatible, but I've never seen them cause issues. Usually it's cuda or other gpu stuff, or conflicts in less regularly maintained packages
- prpl 4y agoI had Stable Diffusion running on m1 and intel macbooks vein the first few days, but the original repo would have done people some favors if they either created proper conda lock files for several platforms or just used conda-forge instead of mixing conda and unnecessarily (I think there was one dep which actually wasn’t on conda-forge, besides their own things) (and actually made the code independent of cuda)
- teaearlgraycold 4y agoAnyone have an M1 Ultra they can test this on? My 3080 Ti can render a 512x512 image in something like 7 seconds and I've love to compare against Apple Silicon.
- latchkey 4y agoM1 max 16" with 64gigs ram, lstein fork, about 30-40s.
- machinekob 4y agoThere isn't any optimised diffuser on m1 yet most of them are just running basic MPS graph or even mix of CPU and MPS ops and ofc it is extremely slow I dont have time to test this implementation but author just use some other implementation with simple UI. So I would be surprise if it is faster than 10s per img with 20steps and probably more close to 25-40s and around 40s-1min per classic 512x512 50-60 step setting as are other models.
- geerlingguy 4y agoOn M1 Max Mac Studio I'm getting about 45s. On my 3080 Ti about 5-7s.
- StrangeDoctor 4y ago
- Bolkan 4y agoClickbait. I clicked on this link. Nothing happened.
- johnklos 4y agoPerhaps "Stable Diffusion on your ARM Mac in 1 click" would've been a more helpful title.
- hunkins 4y agoLove it. If you don't have an M1 Mac, or don't want to wait, https://mage.space https://mage.space does unlimited generations currently. (Note: I am the creator!)
- MuffinFlavored 4y agoI feel like this can't be cheap to run?
- hunkins 4y agoWe're using GPU serverless (via banana.dev), so it's actually not bad. Will have limits at some point, for now go wild.
- stavros 4y agoThanks for the shout, I've made something similar to yours (https://phantasmagoria.stavros.io/ https://phantasmagoria.stavros.io/) and I needed a GPU backend. Trying out their sample script, it seems to take a minute or so to just error out with "taskID doesn't exist" or similar. Have you hit that issue too?
- googlryas 4y agoCan you give an estimate of what the cost is per run?
- fermentation 4y agoThis is really cool and a fun way to try out this stuff I've been hearing about. One thing that'd be cool is a "retry" button that picks a different seed. My first attempt didn't turn out so great (https://i.imgur.com/zV48hCV.png https://i.imgur.com/zV48hCV.png)
- GaggiX 4y agoThe steps size is too low.
- yummybear 4y agoI have a mac a few years old, and now we start seeing M1 only software. My next computer won’t be a mac.
- theodric 4y agoHave they fixed it? I installed it in a virgin macOS 12.5 instance on Wednesday and it didn't work at all
- takoid 4y agoDoes a "1 click" Windows implementation of Stable Diffusion exist yet?
- filoleg 4y agoHad been available for a while, check out NMKD[0]. That's what I've been personally using the entire time. 0. https://nmkd.itch.io/t2i-gui https://nmkd.itch.io/t2i-gui
- ISL 4y agoSeeing copyrighted/trademarked icons in the examples (Darth Vader, for example) really makes me wonder how these models are going to play out in the future. Today, these models are far ahead of the trademark attorneys, but there are powerful interests that are going to want to litigate the inclusion of these entities in the trained models themselves.
- visarga 4y agoI hope people realise these are general purpose text to image systems, not just AI artists. They can be used to generate images for educational content, to generate ideas for design of clothes, shoes, mugs, interior and exterior design, caricatures and memes for social networks, virtual dressing booth, hairstyle and make up, customise games and game characters, they can help create bias detection benchmarks, maybe in the future even generate technical drawings. So the art copyright angle should not be the only one taken into consideration.
- dqpb 4y agoIt's awesome to see how much creativity, progress, and community involvement results from truly open AI development. Congrats to the stable diffusion team for their openness and inclusiveness!
- dharma1 4y agoHow long until on-device stable diffusion on new iOS devices? RAM will be a bottleneck I guess
- russellbeattie 4y agoI downloaded this and tried out a few prompts like, "Mark Twain holding an iPhone", and got back an image of Mark Twain - once in some surrealist nightmare fashion and another more like a 3D render. Neither were holding anything, let alone an iPhone. Cranking up the DDIM slider didn't seem to do much. Trying the same prompt on mage.space (see the creators comment in this thread) produced exactly what I assumed it would. Is there a trick to it?
- pteraspidomorph 4y agoSometimes it's difficult to get certain combinations of things in a picture. It's easier if you provide a basic sketch with the components you want and use img2img (not sure how charl-e has it set up for img2img access, since I use the python original). There's also a knack for writing the prompts, generally you want to write your prompt as a list of short sentences. Don't make your prompt too short. Use concrete and clear concepts, beginning with your main subject, and describe their relation. You can also qualify the background, the mood, the material and the style among other things. Generate more than one sample - I normally go for five or ten samples so there are better odds of getting one that works, since the AI could try to interpret the prompt in different ways that make sense (but not to a human). I can't try it right now, but you might try something like: "A man is standing on the street, close up. The man has an iphone in his hand. The man is Mark Twain. Regular city background. Impressionist. Detailed." Tweak a few times and more often than not you'll end up with something satisfying.
- whywhywhywhy 4y agoYou need to be more specific or you're gonna get a grab bag of 3D renders, stock photos, cartoons, etc. Like what were you hoping for? and add terms that will drive towards that.
- russellbeattie 4y agoI wrote what I was expecting: The same type of results from the online tool I used. My comment was pretty clear, not sure why the two responses I got decided to give me advice about prompts.
- Dotnaught 4y agoReal-ESRGAN and GFPGAN seem to be missing. Those help the image quality significantly.
- offsky 4y agoCan someone please explain how I can run this on my computer but something like GPT3 is too computational intensive to do the same? Isn’t text easier than images?
- WrtCdEvrydy 4y agoGPT3 is a large model that won't fit in memory generally. Stable Difussion isn't so heavy... mostly you are limited by how many steps you want to do.
- d13 4y agoBut why does GPT3 need to be so large?
- yazaddaruvala 4y agoDisclaimer: I’m not well versed in this field, but my basic understanding is that Diffusion can and is “fuzzy”. An image without “crisp” pixels that create lines is acceptable. A basketball with an extra/missing blue pixel is acceptable. With words, they either haven’t invented Diffusion yet, or the nature of the problem is too hard for Diffusion. With text you can’t have the model Diffuse to “teli ne abouf a fdog?” “The dod is graen and has stix leg5”. It’s just too obviously wrong. Meanwhile, the “fuzziness” allows Diffusion models to be smaller, compared with models that need precision.
- Angostura 4y agoLove the 'we haven't managed to implement the ever so complex version checker logic yet - so give us your e-mail' ruse. EDIT: I take it back - all the menus are the generic Electron ones, so it is quite possible that the author is finding this part tricky.
- yamtaddle 4y agoWhen I hit "generate" on my M1 Air with the default options, it just sits saying "initializing...0%" forever. Gave it five minutes, still nothing. Tried twice, same thing. Is it... doing anything? Do I just need to wait 10 minutes? 20?