Big Ticket Items #1195
Replies: 9 comments 17 replies
-
|
While I am now an nVidia user, I was sold on this project when I saw that Vlad fully intended to integrate AMD support (then I used rx590, still secondary) as best he could, which he did, I think he's done the same for Intel's ARC and M1s as best as he could as well. He didn't just ignore it and leave that to be handled by a fork. I feel it would be a bit too... self-interested to focus on one particular GPU manufacturer, even if (or because) that is what the majority are using. |
Beta Was this translation helpful? Give feedback.
-
|
I want my AMD brothers right beside me making art! There shouldn't be any second-class citizens when it comes to AI artwork. |
Beta Was this translation helpful? Give feedback.
-
|
I whole heartily agree with you, while faster gens on the 12 or so cards that would benefit from trt is technically cool, I think supporting other models would be MUCH better for this fork as a whole, it's not like the idea is you get one or the other either, it's just which comes first. I think raw speed is generally less useful for most people, but I don't think a majority of users see it that way, which kinda sucks. |
Beta Was this translation helpful? Give feedback.
-
|
Adding new models seems to me the most difficult task. Without fine tuning, they are all inferior to stable diffusion 1.5. This means that they will still need to add a way to train them, let me remind you that even the version of stable diffusion 2.1 is still not used by anyone. Also optimize, for example DeepFloyd requires 16gb VRAM. Still it will be necessary to maintain all this hodgepodge in working order. |
Beta Was this translation helpful? Give feedback.
-
|
The limitations for using alternative models is the same if not worse than supporting tensor cores, only a handful of devices support them and they can't use any of the fine tuning techniques we use in 1.5, I get wanting moar speed, but I wholeheartedly disagree that it's more logical than starting the groundwork for supporting alternative models that everyone can benefit from. |
Beta Was this translation helpful? Give feedback.
-
no, any tensor cores work would be very specific to existing sd 1.5/2.1 implementation. there is zero chance that would be applicable to other models. |
Beta Was this translation helpful? Give feedback.
-
|
Will backwards compatibility be available in both cases?
|
Beta Was this translation helpful? Give feedback.
-
|
of course! |
Beta Was this translation helpful? Give feedback.
-
|
While I'd like to have the speed of TensorRT, I think that implementing alternative model architectures will help draw attention to them and hopefully bring improvements and optimizations like those that SD gained from the community's work. |
Beta Was this translation helpful? Give feedback.
-
|
I would LOVE Kandinsky model support, it provide really interesting images, also 2.2 version coming soon... |
Beta Was this translation helpful? Give feedback.
-
|
Intel ARC have huge tensor cores |
Beta Was this translation helpful? Give feedback.
-
|
that would be up to intel ipex engine. |
Beta Was this translation helpful? Give feedback.
-
|
Intel ARC 750 - 448 tensor cores Please add tensor cores support for intel ARC!!! |
Beta Was this translation helpful? Give feedback.
-
|
You can't compare Tensor cores from different architectures. |
Beta Was this translation helpful? Give feedback.
-
|
我觉得 张量提升速度和 兼容 ，AUTOMATIC1111和nv已经在做了，等他们实现了再做也不晚，都去做的话浪费人力，而且只能在30系40系的nv卡上应用。 |
Beta Was this translation helpful? Give feedback.
-
|
Microsoft Olive适用于20xx系列的显卡，我在这里找到了这个评论： 我不确定这是否也适用于TensorRT。但我认为应该是的。是的，如果两个开发者现在都在寻找解决方案，然后Nvidia发布了一些内容，那将是很糟糕的情况。但我几乎等不及将TensorRT整合到我的工作流程中了。 |
Beta Was this translation helpful? Give feedback.
-
|
guys, i'd love to know what you're talking about - especially since this is the thread i've started? if you're commenting on a thread that is written in english, comment in english. otherwise, feel free to start a new thread in any language of your choice. |
Beta Was this translation helpful? Give feedback.
-
|
TensorRT or riot <3 |
Beta Was this translation helpful? Give feedback.
-
|
Bring on the riots, daddy wants a new set of models! |
Beta Was this translation helpful? Give feedback.
-
|
For me, as a macOS M1 user, I can only vote for the second option. Haha, it's partly because it brings a new experience and creativity with the new models, and partly because option two utilizes all GPUs. I also have a request: there is really limited optimization and research about running SD on Mac available on the internet, and I don't possess that technical knowledge. I can only hope for technical experts like you, Vlado, to provide insights and guidance. |
Beta Was this translation helpful? Give feedback.
-
|
vlad can do both, he is just being lazy :P |
Beta Was this translation helpful? Give feedback.
-
|
Hahaha, I can only say Vlado, keep it up!!! I need you!!! Especially for those of us who have only 8GB on our Macs, we deeply regret not getting the 16GB option... |
Beta Was this translation helpful? Give feedback.
-
|
I made sure as hell to get 16gb unified for my m2 |
Beta Was this translation helpful? Give feedback.
-
|
Oh, I deeply regret getting only 8GB for my M2. |
Beta Was this translation helpful? Give feedback.
-
|
When I enable four LORA, adetailer, controlnet, and tile - in SD, it runs at an "astonishing speed." |
Beta Was this translation helpful? Give feedback.
-
From two big ticket items on my plate, which one would you like to see more?
For example, TensorRT or any other newer methods like AITemplate, etc.
Note this is applicable to new nVidia GPUs only (3xxx and above)
For example, DeepFloyd, Kandinsky, etc.
Note this is applicable to all GPUs
Note that in both cases, existing workflow will be limited.
For example
Eventually, I hope all is going to get done, but need to prioritize...
179 votes ·
Beta Was this translation helpful? Give feedback.
All reactions