Software Engineering Courses (SECourses)•2y ago

given the announcement page had 15 billion instances of AI safety, i'm thinking SD3 is going to be h

given the announcement page had 15 billion instances of AI safety, i'm thinking SD3 is going to be heavily crippled for anything more complex than a person standing there

Dazzastrous [4090]•2/29/24, 2:15 AM

yeh unfortunatly thats the way things seem to be going, the more this tech improves at such a rate the more people in high places are taking notice

Dazzastrous [4090]•2/29/24, 2:17 AM

ideogram seems to be (at the moment) not that heavily censored

FFabricatedgirls given the announcement page had 15 billion instances of AI safety, i'm thinking ...

Alexm•2/29/24, 2:17 AM

SDXL is the same, if promp following and the architecture is good we can let it learn

AAlexm SDXL is the same, if promp following and the architecture is good we can let it ...

FabricatedgirlsOP•2/29/24, 2:17 AM

training on SDXL was cripped badly

FabricatedgirlsOP•2/29/24, 2:18 AM

i get an error if i try to do stage 2 upscale in supir without doing stage 1 first

Alexm•2/29/24, 2:19 AM

you need to upload to stage 1 image

Alexm•2/29/24, 2:19 AM

your base image

FabricatedgirlsOP•2/29/24, 2:20 AM

FabricatedgirlsOP•2/29/24, 2:22 AM

interesting, even with an upscale value of 1 (and default settings otherwise/no prompt) putting the image in stage 2 and running stage 2 sharpens and cleans the image up a bit

Alexm•2/29/24, 2:24 AM

yea

Alexm•2/29/24, 2:25 AM

and if you smooth it out first, the gun would start looking strange

FabricatedgirlsOP•2/29/24, 2:25 AM

so looks like you can use it to denoise and add detail at same res

Dazzastrous [4090]•2/29/24, 2:28 AM

yeh the metal pops , little extra detail nice one

FabricatedgirlsOP•2/29/24, 2:33 AM

be canceling my magnific subscription soon then lol

Alexm•2/29/24, 2:37 AM

i was never realy a fan of magnific. It kept adding details i didnt want haha

FabricatedgirlsOP•2/29/24, 2:37 AM

yes but sometimes thats a good thing, it can repair faces in SDXL renders

FabricatedgirlsOP•2/29/24, 2:37 AM

and for BG characters

FabricatedgirlsOP•2/29/24, 2:38 AM

so the thing you dont like is the only reason to still use it lol

Alexm•2/29/24, 2:39 AM

yea, but i think if you tamper with the settings of supir, you should be able to do the same

FFabricatedgirls interesting, even with an upscale value of 1 (and default settings otherwise/no ...

Furkan Gözükara SECourses•2/29/24, 2:41 AM

ye it works better in some images depends on taste

Alexm•2/29/24, 2:42 AM

i hope supir gets faster in further development. Its currently really slow :/ even on a A6000

Alexm•2/29/24, 2:42 AM

takes around 10 mins to upscale to 3000x3000

FabricatedgirlsOP•2/29/24, 2:47 AM

Admiral Furkan, do any of your videos deal with multi concept checkpoint training and how to tag those images and whether regularisation images are needed for concepts rather than people

GenZero•2/29/24, 3:24 AM

Scripts eror.I can't Running on local URL Please advise me?.I'm a new member. And then started learning the basics. Please give me some advice.

FFurkan Gözükara SECourses ye it works better in some images depends on taste

VAMP•2/29/24, 3:43 AM

Do did you see this EMO thing... Looks wild right? It can be the open source alternative of heygen..but better ...examples looks promising...
https://www.reddit.com/r/StableDiffusion/comments/1b25pbk/emo_emote_portrait_alive_is_based_on_sd_15/

From the StableDiffusion community on Reddit

Explore this post and more from the StableDiffusion community

felixsanz•2/29/24, 3:48 AM

dafuq....

felixsanz•2/29/24, 3:49 AM

it is public?

FabricatedgirlsOP•2/29/24, 3:49 AM

VAMP•2/29/24, 3:52 AM

Not as of now..but it it better than heygen..it captures better facial expressions and micro expressions..lip movement and mouth movements..even the eyes as well

Vito•2/29/24, 5:38 AM

Hello doctor, I am using your kohya configuration and A5000 acceleration instructions. The fineturn speed is 1.78s/it. Is this a normal speed? It takes about 5 hours to fine-tune 35 pictures. My goal is to fine-tune a 10k data set. This Time is scary.

@Dr. Furkan Gözükara

FabricatedgirlsOP•2/29/24, 6:34 AM

curious what kinda images in the dataset

Vito•2/29/24, 6:44 AM

interior design images

FabricatedgirlsOP•2/29/24, 6:45 AM

what are you using to tag?

FabricatedgirlsOP•2/29/24, 6:45 AM

and do you need regulatisation images to train these?

Vito•2/29/24, 6:45 AM

I also observed that the results were overfitted very early

FFabricatedgirls what are you using to tag?

Vito•2/29/24, 6:46 AM

cog

FFabricatedgirls and do you need regulatisation images to train these?

Vito•2/29/24, 6:48 AM

The original resolution has reached 1024

VVito The original resolution has reached 1024

FabricatedgirlsOP•2/29/24, 6:49 AM

not sure i follow?

TonyHalik•2/29/24, 7:19 AM

what about this error:
c:\Supir\SUPIR\venv\lib\site-packages\open_clip\factory.py:213 in create_model │
│ │
│ 210 │ │ │ │ │ f'Pretrained weights ({pretrained}) not found for model {model_name} │
│ 211 │ │ │ │ │ f'Available pretrained tags ({list_pretrained_tags_by_model(model_na │
│ 212 │ │ │ │ logging.warning(error_str) │
│ ❱ 213 │ │ │ │ raise RuntimeError(error_str) │
│ 214 │ │ │ pretrained_loaded = True │
│ 215 │ │ elif has_hf_hub_prefix: │
│ 216 │ │ │ logging.info(f'Loading pretrained {model_name} weights ({pretrained}).') │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
RuntimeError: Pretrained weights (models/open_clip_pytorch_model.bin) not found for model ViT-bigG-14.Available pretrained tags
(['laion2b_s39b_b160k'].

FFabricatedgirls not sure i follow?

Vito•2/29/24, 7:37 AM

Well, non-native English speaker, you know ha ha

FabricatedgirlsOP•2/29/24, 7:37 AM

I figured ha, all good

Vito•2/29/24, 7:43 AM

It seems like you are also interested in fineturn instead of dreambooth, right?

FabricatedgirlsOP•2/29/24, 7:44 AM

tbh i dont know what's the advantage of either, i thought dreambooth was for finetuning!

Vito•2/29/24, 7:45 AM

Well, I'm also in the exploratory stage

57•2/29/24, 10:57 AM

Here for history, congrats @Dr. Furkan Gözükara your post is going virtal on twitter

557 Here for history, congrats @Dr. Furkan Gözükara your post is going virtal on twi...

VAMP•2/29/24, 11:36 AM

which one share the link

Dr.Hsan•2/29/24, 12:50 PM

@Dr. Furkan Gözükara i am surprised no application has multi prompts for SDXL, its an amazing function which gives much more control on the images. I Used it on through SD official API . By sending 2 positive prompts and one negative. If anyone seen this function please let me know

DDr.Hsan @Dr. Furkan Gözükara i am surprised no application has multi prompts for SDXL, i...

felixsanz•2/29/24, 1:52 PM

multi prompts?

felixsanz•2/29/24, 1:52 PM

you mean one prompt for first text encoder and another prompt to the second text encoder?

felixsanz•2/29/24, 1:52 PM

is not that powerfull!

given the announcement page had 15 billion instances of AI safety, i'm thinking SD3 is going to be h

Similar Threads