Notices by guizzy (in exile) (guizzy@shitposter.club), page 9
-
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 16-Aug-2023 03:07:43 JST 
guizzy (in exile)
Mamma mia! -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 16-Aug-2023 00:36:47 JST
guizzy (in exile)
@why @triodug :blobcatflip: -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 15-Aug-2023 09:49:03 JST
guizzy (in exile)
@why @coolboymew I'm gonna guess it's psychological, the middle space makes everyone feel like they came closer to winning. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 15-Aug-2023 05:25:12 JST
guizzy (in exile)
@lain Not true! -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 15-Aug-2023 03:08:12 JST
guizzy (in exile)
@mangeurdenuage If they were planning to put it on Steam they might hit a little snag there.
But since I've also be working on the problem of computer-generated Stable Diffusion prompts, I am interested in looking at some of their ideas though. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Monday, 14-Aug-2023 22:30:22 JST
guizzy (in exile)
@lain It's easy to see why he's not part of the Chocolate Diplomatic Corps. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Monday, 14-Aug-2023 06:09:41 JST
guizzy (in exile)
Dynavision is pretty good for 3D rendered cartoon style images. I'm still looking for a realistic model. Rundiffusion XL is not bad, but it has a severe case of "same face".
With SD I settled on using 1 2D drawing model, 1 3D render model, 1 realistic model and 1 creative/versatile model. The last one is one that regardless of quality of style, had the highest probability of giving me what I wanted for even with a difficult prompt. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Thursday, 10-Aug-2023 02:51:42 JST
guizzy (in exile)
@Moon @Christmas_Man @meso This is the guy making 4bit quantized models for home use: https://huggingface.co/TheBloke
GPTQ models are for GPU based inference, GGML are for CPU based inference (though you can get speed boost from moving some of the load on your GPU).
I'd recommend installing Oobabooga's Text Generation Webui. It's the Automatic1111 of LLMs: https://github.com/oobabooga/text-generation-webui
With 24Gb VRAM, you can run GPTQ 13b to 20b models with room to spare for extended (over 2048 token) context and keeping Stable Diffusion loaded at the same time. Or you are supposed to be just about able to run 30b models with 2048 context on a headless linux machine. Expect double digit tokens per second. Answers will pop up in seconds.
With GGML models your RAM is going to be your limit, and speed is going to depend on CPU, GPU, RAM speed and how much you can offload to GPU/VRAM. But in general it's likely to be MUCH slower than GPTQ, but if you're running as big a model as you can fit in your machine, expect single digit tokens per second. Expect to wait sometimes over a minute for an answer. Sometimes it's worth it, sometimes not. I've heard people say that the returns from 30b to 70b are quite a bit diminished (ie: it's not really noticeably smarter, just different). -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 23:45:36 JST
guizzy (in exile)
Looking at airline reviews. Air Canada: When we arrived in the plane, there was blood all over my child's seat and seatbelt.
Air Canada sent my teen to the US, stranded him and then refuses to speak to us.
:cate: -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 23:45:36 JST
guizzy (in exile)
Port wine? Cheese platters?
Not for us economy class peasants -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 08:44:21 JST
guizzy (in exile)
Is Stable Diffusion haram if it can be asked to make a picture of Muhammad and complies? -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:47:47 JST
guizzy (in exile)
@waltercool @coin @lain Yeah, but thankfully SDXL can give pretty good results without LORAs -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:38:59 JST
guizzy (in exile)
We need a federated Miiverse alternative. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:30:37 JST
guizzy (in exile)
@lain :sadcat: -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:23:14 JST
guizzy (in exile)
@waltercool @coin @lain Yeah, the Regional Prompter technically blows it all open, but it's tough to use.
I think they recently added SDXL support, I gotta try that. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:07:03 JST
guizzy (in exile)
@piggo @lain I run it with same number of steps as original pass (set to 0 steps), and the choice of upscaler matters. Latent isn't good. I like 4x Valar and I know a lot of people use 4x UltraSharp. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 05:58:05 JST
guizzy (in exile)
@lain @coin Yep, we're only just starting to see some finetunes getting merged for SDXL. The best 1.5 checkpoints by now are a convoluted tree of finetunes and merges, so we're still in the very early stages of building that for SDXL. https://civitai.com/models/124419?modelVersionId=136753
As of now, I find the most interesting thing to do is check how SDXL fares with things that SD 1.5 had a hard time doing. Like raccoons :blobraccoon: -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 08-Aug-2023 01:22:24 JST
guizzy (in exile)
@lain @i @mangeurdenuage -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 08-Aug-2023 01:19:12 JST
guizzy (in exile)
As it always does, attempting to update my Pleroma instance broke it. I'm not sure it's fixable this time. -
guizzy (in exile) (guizzy@shitposter.club)'s status on Tuesday, 08-Aug-2023 01:19:11 JST
guizzy (in exile)
@mangeurdenuage I don't really have any motivation to these days. I'll probably just stay of SPC for a while.
All ０７６萌ＳＮＳ content and data are available under the