Edit model card

์ƒˆ๋กœ์šด SDXL๋ฒ„์ „์ด ๋‚˜์™”์Šต๋‹ˆ๋‹ค.

jjabhongdo1

ํ•œ๊ตญ ์ˆ˜๋ฌตํ™” ๋ชจ๋ธ ์‚ฌ์šฉ ๊ฐ€์ด๋“œ sd1.5๊ธฐ๋ฐ˜

Guide for Using Korean Sumukhwa Model based on SD1.5

์ด์ „์— ์ œ์ž‘ํ–ˆ๋˜ ํ•˜์ดํผ๋„ทํŠธ์›Œํฌ์— ์ด์–ด์„œ ์ด๋ฒˆ์—” ํŒŒ์ธํŠœ๋‹ ํ•œ ๋‚˜์˜จ ๋ชจ๋ธ์„ ๊ณต์œ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋ชจ๋ธ ์‚ฌ์šฉ ๊ฐ€์ด๋“œ๋กœ ๋‚จ๊น๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋Š” ๊ณต์œ ๋งˆ๋‹น์— ์žˆ๋Š” ๊น€ํ™๋„ ๊ทธ๋ฆผ ์ค‘ ์„ ๋ณ„ ํ•œ ์ž๋ฃŒ์™€ Aiํ—ˆ๋ธŒ์— ์˜ฌ๋ผ์™€ ์žˆ๋Š” ํ•œ๊ตญํ™” ๋ฐ์ดํ„ฐ์…‹ ์ž…๋‹ˆ๋‹ค.

Continuing from the previous Hypernetworks, I am sharing the fine-tuned model and providing a guide for its usage. The data used for training includes selected materials from Kim Hong-do's paintings available on the Gongu Sharing Market and the Korean painting dataset on AI Hub.

๋ชจ๋ธ์— ์‚ฌ์šฉ๋œ ์ž๋ฃŒ ์ถœ์ฒ˜

Sources of the data used in the model:

ํ•™์Šต๊ณผ์ •

Training Process

์ด๋ฏธ์ง€๋ฅผ 768ร—768์‚ฌ์ด์ฆˆ๋กœ ๋ฐ”๊พผ๋’ค, clip_interrogator๋ฅผ ํ†ตํ•ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ํ•œ๊ตญ ์ˆ˜๋ฌตํ™” ์ž๋ฃŒ์— gksrnrghk๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ถ™์ด๊ณ , ๊น€ํ™๋„ ๊ทธ๋ฆผ์—๋Š” rlaghdeh๋ผ๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ถ”๊ฐ€๋กœ ๋ถ™์˜€์Šต๋‹ˆ๋‹ค. ์ด ์ด๋ฏธ์ง€๋ฅผ ๋‹ค์‹œ 512ร—512์‚ฌ์ด์ฆˆ๋กœ ๋ฐ”๊พผ ๋’ค Stable Tuner๋ฅผ ์ด์šฉํ•ด์„œ ํ•™์Šต์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์šฉํ•œ ์„ค์ •์€ ์•„๋ž˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.

The images were resized to 768ร—768, and prompts were created using the clip_interrogator. The prompt "gksrnrghk" was added for Korean ink wash painting data, and an additional prompt "rlaghdeh" was added for Kim Hong-do's paintings. These images were then resized to 512ร—512, and training was performed using the Stable Tuner. The following settings were used

  • pretrained model: runwayml/stable-diffusion-v1-5
  • seed: 3434554
  • resolution: 512
  • train batch size: 24
  • num train epochs: 60
  • learning rate: 5e-6

์›๋ž˜ 768 ๋ชจ๋ธ์„ ์ƒ๊ฐํ•˜๊ณ  ์ œ์ž‘ํ–ˆ์œผ๋‚˜, ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ํ™˜๊ฒฝ์ด ์ œ์•ฝ์ด ํฌ๊ณ  ์ปจํŠธ๋กค๋„ท์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์—ˆ๊ธฐ์— 512๋ชจ๋ธ 1.5๋ฒ„์ „์œผ๋กœ ๋‹ค์‹œ ์ž‘์—…์„ ํ–ˆ์Šต๋‹ˆ๋‹ค.

Originally, I intended to create a 768 model, but due to constraints on the execution environment and the unavailability of the Controlnet, I had to work with the 512 model version 1.5.

์ž๋ฃŒ๊ฐ€ ์ด๋ฏธ ์ค€๋น„ ๋˜์–ด์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ฒฝ์šฐ์— ๋”ฐ๋ผ์„œ๋Š” (์˜ˆ์‚ฐ์ด๋ผ๋“ ์ง€) ์ƒˆ๋กญ๊ฒŒ ํ•™์Šต ํ•  ์ง€๋„ ๋ชจ๋ฅด๊ฒ ์Šต๋‹ˆ๋‹ค.

Since the data is already prepared, it may not be necessary to train again in some cases (e.g., budget constraints).

์‚ฌ์šฉ ๊ฐ€์ด๋“œ

Usage Guide

ํ•œ๊ตญ ์ˆ˜๋ฌตํ™” ๋ฐ์ดํ„ฐ๋Š” 6000์žฅ ์ •๋„ ๊น€ํ™๋„ ๊ทธ๋ฆผ์€ 1000์žฅ ์ •๋„ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ์ด์œ ์ธ์ง€, ํ•œ๊ตญ ์ˆ˜๋ฌตํ™” ์Šคํƒ€์ผ๋กœ ํ•˜๋ ค๋ฉด CFG Scale๋ฅผ 2-7 ์‚ฌ์ด๋กœ ๊น€ํ™๋„ ๊ทธ๋ฆผ์˜ ์Šคํƒ€์ผ๋กœ ํ•˜๋ ค๋ฉด 4-12์‚ฌ์ด๋ฅผ ์ถ”์ฒœํ•ฉ๋‹ˆ๋‹ค. ๋‘๊ฐœ์˜ ์Šคํƒ€์ผ ๋ชจ๋‘ ํ™œ์šฉํ•  ๊ฒฝ์šฐ ์ค‘๊ฐ„ ๊ฐ’์œผ๋กœ ํ•˜๋Š” ๊ฒƒ์„ ์ถ”์ฒœํ•ฉ๋‹ˆ๋‹ค. ์Šคํ…์ˆ˜์—๋„ ์˜ํ–ฅ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์ ์ ˆํ•œ ๊ฐ’์„ ์ฐพ๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.

For Korean ink wash painting style, it is recommended to use CFG Scale between 2-7. For Kim Hong-do's painting style, a range of 4-12 is recommended. If you want to utilize both styles, it is recommended to use an intermediate value. The step count also affects the output, so finding an appropriate value is important.

์ž‘๋™์„ ์ž˜ ํ•˜์ง€์•Š์ง€๋งŒ ๊ธฐ๋ฒ•์— ๋Œ€ํ•œ ํ”„๋กฌํ”„ํŠธ๋„ ์ ์šฉ์ด ๋˜์–ด์žˆ์œผ๋ฉฐ ์‚ฌ์šฉํ•  ๋•Œ๋Š” ์•„๋ž˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค. (ํ•˜์ง€๋งŒ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๋งŒ์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค.)

Although it does not perform well, the prompts for the techniques are applied, and when using it, you can use the following prompts. (However, only subtle differences may occur.) Translate it into English.

  • ๋ฐฑ๋ฌต๋ฒ•/Baekmukbeob: baegmyobeob
  • ๋ชฐ๊ณจ๋ฒ•/molgolbeob: molgolbeob
  • ๊ตฌ๋ฅต๋ฒ•/guleugbeob: guleugbeob

๊น€ํ™๋„ ๊ทธ๋ฆผ์„ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์œผ๋ฉด rlaghdeh style, rlaghdeh painting์ด๋ž€ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ฐ™์ด ์‚ฌ์šฉํ•˜๋ฉด ์ข€๋” ๊ฐ•์กฐ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.

If you want to emphasize Kim Hong-do's painting style, using the prompts "rlaghdeh style" and "rlaghdeh painting" together will enhance the emphasis.

์ƒ˜ํ”Œ ์ด๋ฏธ์ง€

Sample Images

txt2img์˜ ์ƒ˜ํ”Œ์ด๋ฏธ์ง€ ์ž…๋‹ˆ๋‹ค.

Here is a sample image from txt2img

jjabhongdo1

gksrnrghk, sky, tree Steps: 40, Sampler: DPM++ 2M SDE Karras, CFG scale: 2.0, Seed: 1271864954, Size: 768ร—512, Model hash: a710c70889, Model: gksrnrghk_15_512_60, Clip skip: 2, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: โ€œgksrnrghk,\โ€gksrnrghk, rlaghdeh\โ€,rlaghdeh โ€œ, Y Type: CFG Scale, Y Values: โ€œ2,3,4,5,7,9,12,15โ€, Version: v1.3.0

ํ•œ๊ตญํ™”์˜ ๊ฒฝ์šฐ ๋†’์€ CFG์—์„œ ํ‘๋ฐฑ์ด ์•„๋‹Œ ์ปฌ๋Ÿฌ๊ฐ€ ๋‚˜์˜ค๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค.

In the case of Korean ink wash painting, colors start to appear at higher CFG values instead of black and white.

Downloads last month
162

Space using gagong/korean-sumukhwa-model-ver-1 1