Software Engineering Courses (SECourses)•2y ago

Hello doctor, I am using your kohya configuration and A5000 acceleration instructions. The fineturn

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

Fabricatedgirls•2/29/24, 6:34 AM

curious what kinda images in the dataset

VitoOP•2/29/24, 6:44 AM

interior design images

Fabricatedgirls•2/29/24, 6:45 AM

what are you using to tag?

Fabricatedgirls•2/29/24, 6:45 AM

and do you need regulatisation images to train these?

VitoOP•2/29/24, 6:45 AM

I also observed that the results were overfitted very early

FFabricatedgirls what are you using to tag?

VitoOP•2/29/24, 6:46 AM

cog

FFabricatedgirls and do you need regulatisation images to train these?

VitoOP•2/29/24, 6:48 AM

The original resolution has reached 1024

VVito The original resolution has reached 1024

Fabricatedgirls•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?

VitoOP•2/29/24, 7:37 AM

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

Fabricatedgirls•2/29/24, 7:37 AM

I figured ha, all good

VitoOP•2/29/24, 7:43 AM

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

Fabricatedgirls•2/29/24, 7:44 AM

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

VitoOP•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!

Ffelixsanz multi prompts?

Dr.Hsan•2/29/24, 1:56 PM

import fs from "fs";

const options = {
method: "POST",
headers: {
Accept: "application/json",
"content-type": "application/json",
Authorization: "Bearer YOUR_API_KEY"
},
body: JSON.stringify({
samples: 1,
height: 1024,
width: 1024,
steps: 40,
cfg_scale: 5,
textprompts: [
{
"text": "A painting of a cat",
"weight": 1
},
{
"text": "a dog barking",
"weight": 1
},
{
"text": "logo, ",
"weight": -1
}
],
})
};

const response = await fetch(
"https://api.stability.ai/v1/generation/stable-diffusion-xl-1024-v1-0/text-to-image",
options
)

if (!response.ok) {
throw new Error(Non-200 response: ${await response.text()}Non-200 response: ${await response.text()})
}

const responseJSON = await response.json()

responseJSON.artifacts.forEach((image, index) => {
fs.writeFileSync(
`./out/txt2img${image.seed}.png`,
Buffer.from(image.base64, 'base64')
)
});

felixsanz•2/29/24, 1:56 PM

that doesn't answer

Ffelixsanz multi prompts?

Dr.Hsan•2/29/24, 1:56 PM

both prompts are flowing through both encoders

felixsanz•2/29/24, 1:56 PM

but yeah, probably one prompt for each text encoder

DDr.Hsan both prompts are flowing through both encoders

felixsanz•2/29/24, 1:57 PM

that's not possible. it's just a concatenation then

felixsanz•2/29/24, 1:57 PM

it's like sending "A painting of a cat a dog barking"

Ffelixsanz but yeah, probably one prompt for each text encoder

Dr.Hsan•2/29/24, 1:58 PM

you can add more than 4 prompts, control the weight for each 1 seperatly.

felixsanz•2/29/24, 1:59 PM

anyway, all apps have that

felixsanz•2/29/24, 1:59 PM

it's called prompt weighting

Ffelixsanz it's called prompt weighting

Dr.Hsan•2/29/24, 2:00 PM

https://platform.stability.ai/sandbox/multi-prompting have a look here i think you will be able to generate free images

felixsanz•2/29/24, 2:02 PM

(A painting of a cat:0.8) (a dog barking:1.4)

felixsanz•2/29/24, 2:02 PM

that's how you do it in a1111

felixsanz•2/29/24, 2:02 PM

there is also other types of blending

felixsanz•2/29/24, 2:02 PM

it comes from the https://github.com/damian0815/compel library

felixsanz•2/29/24, 2:03 PM

specifically: https://github.com/damian0815/compel?tab=readme-ov-file#120---concatenate-embeddings-using-and

felixsanz•2/29/24, 2:04 PM

The syntax is ("prompt part 1", "prompt part 2").and(). You can have more than one part, and you can also weight them, eg ("a man eating an apple", "sitting on the roof of a car", "high quality, trending on artstation, 8K UHD").and(1, 0.5, 0.5)

Furkan Gözükara SECourses•2/29/24, 2:06 PM

v14 https://www.patreon.com/posts/99176057

Furkan Gözükara SECourses•2/29/24, 2:06 PM

felixsanz•2/29/24, 2:09 PM

@Dr. Furkan Gözükara I would like to know what your secret is to make videos with so much work and also having time to make those scripts and posts for Patreon