CompVis
Stable-diffusion v-1-5
Hi, amazing work!
I see that the new v-1-5 model is available on dream studio, will the weights be released as well? Can we also have details of what was changed in the training?
Thank you
I'm only seeing v1. Am I missing something? If v1.5 is about to come out I'd expect to see v1.4.
I'm only seeing v1. Am I missing something? If v1.5 is about to come out I'd expect to see v1.4.
1.4 weights available since the beginning
You can click on the 'subscribe' button in the sidebar to subscribe to an issue without notifying everyone that you're subscribing to the issue.
My best guess for a release date: 25th
Reasoning: https://twitter.com/EMostaque/status/1567744932625915912
This is a very interesting question, why would they give 1.5 away? why the free lunch? what's the cost of training and who/what is paying for this?
This is a very interesting question, why would they give 1.5 away? why the free lunch? what's the cost of training and who/what is paying for this?
like there is a huge difference between 1.4 and 1.5 i've tested 1.5 weights on the website and it gives results really similar to 1.4 ver
@xornotxor
This is a very interesting question, why would they give 1.5 away? why the free lunch? what's the cost of training and who/what is paying for this?
stability.ai is the group behind stable diffusion.
From their FAQ
We’re a company of builders who care deeply about real-world implications and applications. Many of our most considerable advances grow from working across multiple teams. We are unafraid to go against established norms and explore creativity.
Our primary drive is to generate breakthrough ideas and convert them into solutions. We respect innovation over tradition. We trust that our differences make us more robust, and so we seek reason within every difference of perspective.
Seems like they open their model to get as many eyes on it as possible. IIUC A lot of the optimizations and additional features were discovered / created by the community around SD and eventually rolled into dream studio (stability's commercial product)
Standard open source mentality. Most people would rather just pay for dream studio, but technically inclined folk can run (and eventually contribute to) stable diffusion.
like there is a huge difference between 1.4 and 1.5 i've tested 1.5 weights on the website and it gives results really similar to 1.4 ver
I think this is an interesting phenomenon. For me, the 1.5 weights behave much better, yet I've seen other people say this as well. I think a big part of it is how you're using Stable Diffusion. I try to make prompts that give me varied, creative results, such as "1970s movie poster by Frank Frazetta" or "tarot art card, embrace ordinary omnipresent" (plus a bunch of style prompt) The 1.5 weights seem to give much more variety and much more coherence (for example, better anatomy) given these types of prompts. For someone describing a very specific scene, perhaps there's much less difference.
I think this is an interesting phenomenon. For me, the 1.5 weights behave much better, yet I've seen other people say this as well. I think a big part of it is how you're using Stable Diffusion. I try to make prompts that give me varied, creative results, such as "1970s movie poster by Frank Frazetta" or "tarot art card, embrace ordinary omnipresent" (plus a bunch of style prompt) The 1.5 weights seem to give much more variety and much more coherence (for example, better anatomy) given these types of prompts. For someone describing a very specific scene, perhaps there's much less difference.
well, i decided to test it for example
tarot major arcana hierophant intricate details, artstation trending by mandy jurgens, jacub rosalski, hyperdetailed hyperrealistic tempera
1.4 full ema:
1.5:
can't really see much difference there
Yeah ok, I think you're right. 3 different prompts, all on DreamStudio:
1.4
1.5
1.4
1.5
1.4
1.5
That's me selecting prompts I thought were most different between 1.4 and 1.5. I was running 1.5 in DreamStudio and 1.4 in CoLab. I think the sheer speed of rerolling images in DreamStudio was getting me strikingly better results.
Thanks Ben,
I was referring to the $600,000 cost it took to train 1.4, why would they give 1.5 for free?
Stable Diffusion was trained on a subset of the LAION-Aesthetics V2 dataset. It was trained using 256 Nvidia A100 GPUs at Amazon Web Service for a total of 150,000 GPU-hours, at a cost of $600,000.
@xornotxor
This is a very interesting question, why would they give 1.5 away? why the free lunch? what's the cost of training and who/what is paying for this?
stability.ai is the group behind stable diffusion.
From their FAQ
We’re a company of builders who care deeply about real-world implications and applications. Many of our most considerable advances grow from working across multiple teams. We are unafraid to go against established norms and explore creativity.
Our primary drive is to generate breakthrough ideas and convert them into solutions. We respect innovation over tradition. We trust that our differences make us more robust, and so we seek reason within every difference of perspective.
Seems like they open their model to get as many eyes on it as possible. IIUC A lot of the optimizations and additional features were discovered / created by the community around SD and eventually rolled into dream studio (stability's commercial product)
Standard open source mentality. Most people would rather just pay for dream studio, but technically inclined folk can run (and eventually contribute to) stable diffusion.
coming soon or next year depends on how you interpret it reason: https://twitter.com/EMostaque/status/1574141762750369794?cxt=HHwWhMC-rfrlvNgrAAAA
coming soon or next year depends on how you interpret it reason: https://twitter.com/EMostaque/status/1574141762750369794?cxt=HHwWhMC-rfrlvNgrAAAA
https://en.wikipedia.org/wiki/Rosh_Hashanah
Yeah I don't think anything has been said regarding a release date for 1.5 weights. People in this thread just be guessing
I really hope they do release it, otherwise they would be pretty hypocritical.. not that they have all rights not to do so, but still.. yet another company which consolidates this creative power in the hands of a few individuals behind a paywall...
I have heard it from a very, VERY reliable souce of a friends friend whom heard from their Uncle who works for a company where a coworker has a brother whom knows a guy that has a friend who shares a cubicle next to a girl that likes to read the Stable Diffusion public "Issues" mesages who said that she read a post from a guy who knows a guy that lives in her apartment who saw on a Reddit forum that someone heard a converstaion on an elevator where a Stable Diffusion coder told another guy about needing more brain power so had decided to bring in his brother that works for a different company where a coworker has a cousin whom knows a guy that has a friend who shares a cubicle next to a guy that likes to read the Stable Diffusion public "Issues" mesages who said that he read a post from a guy who knows a guy that lives in his apartment that saw on a Reddit forum that someone heard a converstaion on an elevator where a Stable Diffusion coder told another guy that he spoke with another coder and that they had decided to bring in his sister who works for a company where a coworker has a cousin whom knows a girl that has a friend who shares a cubicle next to anither girl that likes to read the Stable Diffusion public "Issues" mesages who said that she read a post from a guy who knows a guy that lives in her apartment who saw on a Reddit forum that someone heard a converstaion on an elevator where a Stable Diffusion coder told another guy about needing more brain power so had decided to bring in his nephew that works for a different company where a coworker has an in-law whom knows a guy that has a friend who shares a cubicle next to a girl that likes to read the Stable Diffusion public "Issues" mesages who said that she read a post from a guy who knows a guy that lives in her apartment that saw on a Reddit forum that someone heard a converstaion on an elevator where a Stable Diffusion coder told another guy that he spoke with another coder and that they had decided to bring in his.......
Yeah. It's definitely coming..... why? Because I heard it from a guy. 😜
I have heard it from a very, VERY reliable souce of a friends friend whom heard from their Uncle who works for a company where a coworker has a brother whom knows a guy that has a friend who shares a cubicle next to a girl that... [ 20 more lines of this nonsense ] ... that someone heard a converstaion on an elevator where a Stable Diffusion coder told another guy that he spoke with another coder and that they had decided to bring in his.......
Yeah. It's definitely coming..... why? Because I heard it from a guy. 😜
I think this comment could use some refactoring...