Luffy wallpaper 4k

I have been writing a novel for a couple of months, and I'm using stable diffusion to illustrate it. The advent of AI was a catalyst for my imagination and creative side. :)
As so many others in similar situations, a recurring problem for me is consistency in my characters. I've tried most common methods, and have, after lots of testing, experimenting and primarily FAILING, now reached a point where I think I have found a good enough workflow.
What I wanted: A method that lets me generate:

1. The same recognizable face each time
2. The same clothing*
3. Able to do many different poses, expressions, angles, lighting conditions
4. Can be placed in any environment

\This appears to be near-impossible. I have settled for “similar enough that it’s not distracting”.*
Here are some examples of the main character in my story, Skatir:
​
Skatir 1
​
Skatir 2
​
Skatir 3
If you are interested on seeing the results of this process applied in practice (orr just listen to an epic fantasy story), check out my youtube page where chapter 1- 3 is currently up: https://www.youtube.com/playlist?list=PLJEcSn1wDRZsGuSBa87ehc7-VWYQNraIt
​
My process can be summarized into the following steps:

1. Generate rough starting images of the character from different angles
2. Detailed training images, img2img of ~15 full-body shots and ~15 head shots
3. Train two Loras, one for clothing and one for face
4. Usage the two Loras together, one after the other with img2img

Detailed description of each step below
​
Step 1. Rough starting images
Generate a starting image with charTurner [1]. You want the same clothing in 3-4 different angles. Img2img with high denoising can help create the desired number of angles. See example below.

1. CharTurner is a bit sensitive with what model you use it with. I’ve had decent results with DreamlikeArt [2]. Note that these images are just for creating a very rough base, and that exact style and amount of details does not matter here.
2. In principle any method could be used to get these starting images. The important thing is that we same clothes and body type.

Starting image for charTurner. USe this as init image with denoising ~0.8
Output from lots and lots of runs with charTurner.
Step 2. Detailed training images
Next step is to split the output image into at least 30 images (15+15), in the following way:

1. Full-body portraits and half-shots (waist up) portraits for each angle
2. Head close-ups. Varying levels of zoom angles.

Then add details to each image using img2img on each image.
A: For full-body and half-shots;

1. Decide what you want, and rerun img2img until you get what you want.
2. For each image, alter details such as lighting.
3. Use comprehensive and descriptive prompts for clothing.
4. Denoising strength 0.3 - 0.5.
5. Use neutral backgrounds

​
Fullbody images after img2img for more details
​
Example of fullbody image after img2img for more details
B: For head close-ups,

1. Use loras or embeddings to add consistency and detail. I have used multiple embedding of real people. It keeps results consistent but ensures that end result doesn’t look too much like any one single specific person.
2. Denoising strength 0.3 - 0.5.
3. For each image, alter details such as lighting, facial expression, mood.
4. Use neutral backgrounds

Face images after img2img for more details and expressions
​
Example of face closeup after img2img for more details and expressions
Step 3. Train Loras
TBH I am kind of lost when it comes to actual knowledge on Lora-training. So take what I say here with a grain of salt. What I have done is:
A: Train two Loras. I've found that this approach with two loras vastly improves quality.

1. LoraA dedicated to clothing and body type, and
2. LoraB dedicated to the head (face and hair).

B: Tagging images I have found does not make much of a difference in end results, and sometimes makes it worse. I am using extremely simple tagging:

1. "full-body portrait of woman" and
2. "Close-up portrait of woman".

For Lora-settings, I am just running with the default settings in kohya-trainer [3], and Google colab since my computer is not good enough for training. Anylora [4] as base model (this of course depends on what model you want to use later). I'm mostly using revAnimated [5] or similar models, which works okay with AnyLora.
Step 4. Usage the two Loras together
There are three steps to this. In some cases you can jump straight to step 2 or 3, depending on how complicated images you want. E.g. if I only want a closeup on the face, I go directly to step 3.

1. General composition
   1. Start without a Lora at all.
   2. Prompt for background
   3. Describe your character in very generic terms (I use “ginger girl in black dress”)
   4. Re-run until you get decent results
   5. Adjust character clothing and hair in image editing software (I use GIMP)
   6. Upscale. I use img2img with the same prompt but bigger resolution to upscale
2. Body
   1. Use the body Lora
   2. Img2img or inpainting from general composition image. Denoising strength 0.4 - 0.5.
   3. Prompting. Use a standard structure to improve consistency. For me, that's the parts about clothing and hair. Add background, pose, camera orientation. Prompt could look something like this:
      1. , a portrait of a young woman, teen ginger girl, short bob cut, ginger, black leather dress, brown leather boots, grieves, belt around waist, fantasy art, 4K resolution, unreal engine, high resolution wallpaper, sharp focus
   4. As with all AI-art where you are after something specific, be prepared to do multiple iterations, and use inpainting to fix various details, etc.
3. Face
   1. Use the head lora.
   2. Img2img or inpainting on the image where you have body correct. Denoising strength 0.3 - 0.4.
   3. Prompting. Again use a standard structure to improve consistency. For me, that's the parts about hair, eyes, age etc. Add facial expression, camera placement, etc. Prompt could look like this:
      1. , large grin, bright sunlight, green background, a portrait of a young petite teen, blue eyes, norse ginger teen, short bob cut, ginger, black winter dress, fantasy art, 4K resolution, unreal engine, high resolution wallpaper, sharp focus

Below is an example of this used in practice.
Step 1: General composition
Prompt: “((best quality)), ((masterpiece)), (detailed), ancient city ruins, white buildings, elf architecture, ginger girl in jumping out of a window, black dress, falling, bright sunlight, fantasy art, 4K resolution, unreal engine, high resolution wallpaper, sharp focus”
(here using the model ReV Animated [4])
Do many attempts and pick one that you like. I like to start with smaller images and only upscale the ones I like. Preferable upscale before moving to next step.
I like the pose and the background in the image marked with green "circle". But some details are too far off from my character to easily transform her to Skatir. E.g. hair is to long, and she has mostly bare arms and legs. I make very simplistic editing in GIMP to adjust for this.
Adjust in image editing software. In this case I made the hair shorter, gave her brown boots and white shirt:
Step 2: inpaint with body lora.
Using inpaint, I tranform the generic girl in the original image to Skatir
Prompt: “, a portrait of a young woman falling, teen ginger girl, short bob cut, jumping out of a window, black leather dress, brown leather boots, grieves, belt around waist, fantasy art, 4K resolution, unreal engine, high resolution wallpaper, sharp focus”
Inpaint with body-Lora
Now this is starting to look like Skatir. Next I use inpainting to fix some minor inconsistencies and details that don't look good. E.g. hands look a bit weird, boots are different, and I don't want any ground under her (in this situation she has jumped out of a window!).
Fix details with more inpainting!
Step 3: Inpaint with head lora.
Final step. Make the face look like the character, and add more detail to it (human attention are naturally drawn to faces, so more details in faces are good). Just inpaint her face with lora + standard prompt.
Prompt: “, scared, looking own, panic, screaming, a portrait of a ginger teen, blue eyes, short bob cut, ginger, black winter dress, fantasy art, 4K resolution, unreal engine, high resolution wallpaper, sharp focus”
​
Final version
There you have it! I hope this helps someone.
Resources:
[1]: charTurner: https://civitai.com/models/3036/charturner-character-turnaround-helper-for-15-and-21
[2]: Dreamlikeart: https://civitai.com/models/1274?modelVersionId=1356
[3]: kohya Lora trainer: https://github.com/Linaqruf/kohya-traineblob/main/kohya-LoRA-dreambooth.ipynb
[4]: ReV Animated https://civitai.com/models/7371?modelVersionId=46846
​
If you have ideas on how to make this workflow better or more efficient, please share in comments!
If you are interested in finding our why this girl is jumping out of window, check out my youtube page where I post my stories (although this takes place in a future chapter that I have not yet recorded).

https://rare-gallery.com/396721-wallpaper-kayn-lol-art-4k-hd.html And should Chase use shen gong wu and use make a challenge in a xiaolin showdown!

Here is the link where I found it:
https://www.heroscreen.cc/2022/02/iphone-wallpaper-4k-colorful-style.html?m=1

​
https://preview.redd.it/o8guq6slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=92bb6fc9ef52a6780cc1b0a0b339aec2af5fa1eb
https://preview.redd.it/87kndaslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=ea175a398eb574ba69a2177bbe353ed949d77b18
https://preview.redd.it/2duqh8slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=5a0c209fd3a61dafc400814084fbb71ec9cedeeb
https://preview.redd.it/sgeuy6slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=cf599c8d117793b15ba49dc1f361ceba0eecf81e
https://preview.redd.it/yud1u6slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=b399cbf1f7ebb39a606e80437d2402a13b545308
https://preview.redd.it/knm7y8slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=18ff619058410a71d3914218e2dc667261f35eeb
https://preview.redd.it/85nqv9slx74b1.jpg?width=3840&format=pjpg&auto=webp&s=e5235f5252a7447c436f9c453854df50f59733e7
https://preview.redd.it/707izcslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=b30b2e243c65cd5309a9bfbb3bba51bb4277c88d
https://preview.redd.it/z7trjaslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=9689ee7879f7869f96b1568c09ff5bd5b6ea9514
https://preview.redd.it/7fvt9jslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=cd2c1a905c6899980acee8e213e8d389bf8f1e78
https://preview.redd.it/x4q4ocslx74b1.jpg?width=2160&format=pjpg&auto=webp&s=9cf890318520c853071539b184eecd767e804b6e
https://preview.redd.it/d79xfdslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=3d745b1a90fb6744d5064294c1cc9f187fcf799c
https://preview.redd.it/vuj583tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=68dc0b56eaf08ce5ce2c4384c1d7b613dd0d65ee
https://preview.redd.it/p4ozl3tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=cd55055ded29d794b3792268cad9ef44a837d15f
https://preview.redd.it/cvyyxfslx74b1.jpg?width=3840&format=pjpg&auto=webp&s=2001b79dd1ea1b70a7aa5af3deaaee9e510f570e
https://preview.redd.it/bh9ih6tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=bcf1deea562ce3b08e06cdc8d6f06892bd04ce38
https://preview.redd.it/qaetz2tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=7002639f79d1bce7c9622dd660cf495808c564d1
https://preview.redd.it/qeiwj5tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=6678a357c24efbf0a26a27f22868f4e74d82a83d
https://preview.redd.it/ppb4b6tlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=6c9dd3041f4aeadc6454b22d42fc9633aca5c68b
https://preview.redd.it/4qgztftlx74b1.jpg?width=3840&format=pjpg&auto=webp&s=47db73425c5af2921207fc69425872a84049c5c1