Traditional Korean Painting Model

(Translated from English using DeepL)

I renamed the model because I felt it was creating unnecessary misunderstandings. I fine-tuned it with data from Joseon era painters, modern Korean paintings, and ink paintings.

Hypernetworks and an SD1.5-based model were created a few months prior, and shortly afterward, a model named SDXL was released. At the time, I didn’t work on the SD2.1-based model because I was hesitant to use ControlNet and because it wasn’t clearly superior to SD1.5. While prompt-driven changes are more evident, reproductions of known subjects appear slightly better, yet there was no clear advantage over the higher system requirements. However, with SDXL, it was said that ControlNet and various other features would be supported out of the box, and it did not require excessive specifications for supported resolutions, so I decided to give it a try.

For more information on how it was created, see this article.

HuggingFace Link

https://huggingface.co/gagong/Traditional-Korean-Painting-Model-v2.0

https://huggingface.co/gagong/korean-sumukhwa-model-ver-1

Used data

Joseon Dynasty Painters Link

Since paintings by Joseon-era painters were limited in number, I increased the quantity by fragmenting the images and applying horizontal flipping. Total: 5,684.

Digital K-art data Link

Reduced from 11,246 to 6,253 images. AIHUB contains an enormous amount of data, so I first narrowed it down to labeled data, resulting in 11,246 images. From those, I selected images closest to a square shape, resulting in 6,253 images.

Production data for traditional Korean ink paintings by style Link

Used 5,860 images cropped into squares. This dataset is extremely large, so the quantity could be increased; however, I intentionally matched it to a similar amount as the Joseon-era painter data.

A sample of this data is shown below:

User Guide (SDXL version)

Usage Tips

1. In general, higher CFG scale leads to worse quality. We don’t recommend setting it above 10. However, shorter words appear to be less affected.

2. The longer the sentence you type, the more likely the output will resemble a generic photograph. Increasing the weight of trigger words does not improve the result.

3. The recommended CFG scale is between 4 and 8.

4. Do not use low-quality trigger words in isolation.

5. I haven’t tested using multiple trigger words simultaneously, but I suspect it generally works better.

6. When the image turns out looking too photographic, using prompts like “painting” or “style” may yield better results. For example: kimhongdo painting

7. As this is still an early stage of use, the methods described here may be revised later.

Trigger Words List

Note: This does not accurately reflect the true painting style of the actual painters.

View as a larger image Link

• whtjs
  Emulates the average painting style of a Joseon-era painter.

• kangsehwang
  Drawn in its simplest form.

• kimhongdo
  Highest quality.

• sinyunbok
  Second highest quality.

• simsajeong
  Higher probability of a woman’s face appearing in the image.

• anjungsik
  Lowest quality.

• jangseungeop
  Second worst quality.

• heoryeon
  Produces the result with the least color.

• gksrnr
  Creates modern forms. Body errors are frequent.

• tnanr
  Produces an average of the three ink painting techniques below.

• baengmyo
  Results are often similar to those of ‘whtjs.’

• gureuk
  The result features a picture with a thin border.

• molgol
  Produces results without clear boundaries.

User Guide (SD 1.5 version)

Trigger Words: GKSRNRGHK, RLAGHDEH

Approximately 6,000 images of Korean ink wash paintings were used, and about 1,000 Kim Hong-do-style paintings. Due to this, for Korean ink wash painting style, a CFG Scale of 2–7 is recommended; for Kim Hong-do-style painting, 4–12 is recommended. If using both styles together, an intermediate value is advised. The step count also influences results, so finding an appropriate value is important.

Although performance is not optimal, prompts related to techniques are still applied. When using, you can use the following prompts. (However, only subtle differences will be observed.)

• Baekmukbeob (White Ink Method): baegmyobeob

• Molgolbeob (No Boundary Method): molgolbeob

• Guleugbeob (Fine Line Method): guleugbeob

To emphasize Kim Hong-do’s style, use the prompts “rlaghdeh style” and “rlaghdeh painting” together.

Sample Images

This is a sample image from txt2img.