Dezgo : Detail Instructional Doodling Process Korean Pop Dancers

Finally I see sun light , out from the caves of mental block. See previous blog https://prataverse.blogspot.com/2023/08/dezgo-sculptural-doodle-class.html 

This blog gives a peek into the details of my Instructional Doodling process. It can be an assemblage of my doodling and my bad drawing processes.

 These three personal lexicons that are important :

• Doodle : Subconsciously creating line work with my fingers until I arrive at a satisfied image. The line work usually flows in a continuous manner.
  
  
• Bad Drawing : once I know what the doodle image is going to be means that I am conscious and have a purpose in creating line work to complete the image. However, it is bad because I am not trying to draw the image in correct proportion or to draw a beautiful composition.
  
  
• Drawing : normally a process to correct the image, to redraw for the AI to generate better.

  

K-pop Dancer #1

 

STEP 1 : For this particular drawing , I doodle the eyes, nose and lips , bad drawing the head and the hair.

 

STEP 2 : Title it as Korean Pop Dancer

STEP 3 : I scaled the head smaller and bad drawing the rest of the body in a dance pose. 

STEP 4 : First image completed.  I start the process again to process the next Korean Dancer Image.... until I stop doodling or bad drawing.

 

STEP 5 : Decide what I want for the AI generation to get it accurate. For this case I want the AI to :

• follow the body pose of the input image.
• follow the pointing fingers of the input image
• follow the boots of the input image
  
• follow the text prompt a realistic image of a K PoP dancer
  

 

Control Model : Scribble

Model : AbsoluteReality 1.6 

 

PROMPT : Realistic Rendering Korean male dancer

Control Scale : 69%

Guidance : 14

 

The hands did not turn up well and the clothing looked old school. I would preferred it is modern Korean pop style.

 PROMPT : Realistic Rendering Korean pop male dancer
Control Scale : 85%
Guidance : 14

STEP 6 : Generate image with DEZGO AI . Choosing the best overall generation.

 

Prompt :Pointing white glove hand

 STEP 7 : Correct hands Adobe Photoshop beta AI Generative Fill. White glove is a good idea because it is difficult to match skin tone.

K-pop Dancer#2

 

STEP 1 : For this 2nd drawing , I doodle the eyes, nose and lips , bad drawing the head and the hair.

STEP 2 : Scale the head down. Bad drawing the body but realized that I have no space.

STEP 3 : Scale the half body and rotate to the landscape to have more space to bad drawing complete the whole image.

STEP 4 : Decide what I want for the AI generation to get it accurate. For this case I want the AI to :

•     follow the body pose of the input image.
•     follow the text prompt a realistic image of a K PoP dancer

STEP 5 : Generate the initial Control Text to Image using Dezgo. The hand is not placed on the floor as I have loosen the control setting. The other hand is inward could be due to the Resolution Landscape.

PROMPT : Realistic rendering of Korean pop male dancer doing breakdancing

 

STEP 6 : Regenerate by increasing the height of the image by moving the resolution a notch from extreme landscape.

I prefer the 2nd image so I take it to Adobe Photoshop beta to AI edit.

 

STEP 7 : Extend the image using Adobe Photoshop beta generative AI.

 

 

STEP 8 : The hand is not generated well probably due to beta but more or less the same color tone. Select and regenerate the hand.

 

 

K-pop Dancer #3

STEP 1 : For this 3nd drawing , I doodle the eyes, nose and lips , bad drawing the head and the hair.

 

STEP 2 : I scaled the head smaller and bad drawing the rest of the body in a dance pose. 

STEP 3 : Decide what I want for the AI generation to get it accurate. For this case I want the AI to :

•         follow the body pose of the input image.
•         follow the text prompt a realistic image of a K PoP dancer

STEP 4 : Generate the initial Control Text to Image using Dezgo. There are lines on the extending out from the shoes due to the doodle image. The other hand is not clenched. I forgot to put the "fists" or "hands"  in the prompt.

PROMPT : Realistic rendering of Korean pop male dancer with clenched

 

STEP 5 : I update the doodle again using the drawing process to correct for AI. 

 

• redraw the end of the pants
• redraw the hands.

STEP 6 : Generate the image using the updated drawing. There are lines on the extending out from the shoes due to the doodle image. The other hand is not clenched. I forgot to put the "fists" or "hands"  in the prompt.

 

 PROMPT : Realistic rendering of Korean pop male dancer with clenched fists

Seed : 3169676413

 

I do like the 2nd image generation better so I take down the Seed number.

STEP 7 : Regenerate images using the seed number.

STEP 8 : After evaluating the generated images in Step 6 and 7. I have figured out the bug is actually due to my bad drawing. The negative spaces ( see red arrows ) are interpreted as fists by the AI.  

 

• Redraw the left side negative space. 
• Scale down the head.
  

 

 

 

STEP 9 : Regenerate image using updated drawing. I like that the dancer pose like a puppet. I can add the thumbs in Adobe Photoshop beta but it looks good for me.

 

K-pop Guitarist #4

STEP 1 : For this particular drawing , I doodle the head, the face as well as the body. I realized later that the hands are circular and part of the circular hole of the guitar where the strings are connected to. The other hand seems to hold a pick that I do not remember during doodling. 

STEP 2 : Decide what I want for the AI generation to get it accurate. For this case I want the AI to :

•  follow the hair style from the initial image and prompt 
  
•  follow the text prompt a realistic image of a K PoP guitar
•  produce a decent electric guitar
  

STEP 3 : Generate the image. The hands did not generate properly.

PROMPT : Realistic rendering of Korean pop male guitarist with large hair

STEP 4 : Use Adobe Photoshop beta to edit the hands.

See the next blog : https://prataverse.blogspot.com/2023/08/dezgo-set-up-lora1-from-civitai-dark.html

Dezgo: Sculptural Doodle Class

 After doodling for sometime , I can now classify one type of doodle as Class : Sculptural Doodle. Lets explore what setting of Dezgo is best suited for this class type. 

The stone tablets by Shin Beomsum in Singapore Biennale has provided me the inspiration for Sculptural Doodles type https://prataverse.blogspot.com/2023/01/singapore-biennale-2022-natasha.html

I did a sculptural doodle turning out looking like a pupa. Probably subconsciously my mind telling me to wait a little longer for transformation out from the period of doodling drought or mental block faced in my previous blog. See https://prataverse.blogspot.com/2023/07/dezgo-image-to-image-other-side-of-color.html    

Lets see what image AI can generate in Controlled Text to Image.

PROMPT : " "

CONTROL MODEL : Depth Map

MODEL : DreamShaper 7

CONTROL SCALE : 85%
GUIDANCE : 14

I leave the prompt empty to let the AI figure out what it is. If I were to prompt "Pupa" it will generate images related to an insect pupa stage of life. I set the Control Model : depth map as I want AI to interpret my doodle as 3D with depth. I set the Control Scale 85% and Guidance 14 to want the AI to build images based on my doodle. 

 

The results were all from the same setting are impressive, ranging from massive alien form to small size science fiction product design. 

The skeptical engineering mind question about the repeatability of the setting applied to other class sculptural doodle.

Lets do another sculptural doodle that did not generate well for Image to Image.

 

  

It interpreted my doodle quite accurately stairs leading up a cliff up a minimalist architectural structure. 

 One of the interesting interpretation is a Japanese girl with a strange hat.

I know that in the above AI generation the AI is applying the concepts into the initial image that I have provided.

This setting however does not solve all cases. When the AI fail to conceptualize the image , it just present it in a different style. Then maybe adding a prompt might help to steer the AI to the right direction. See the case as follows. 

The Controlled Text to Image does not know what it is , it  merely only refine my line works.  

I think Image to Image did a better job with the prompt "marble sculpture". It has generated the partial face of the person which is concept enough for me.  https://prataverse.blogspot.com/2023/07/dazgo-ai-image-to-image-generation.html

 See Next blog after mental block has been removed. https://prataverse.blogspot.com/2023/08/dezgo-detail-instructional-doodling.html