Drawing Avatars with Stable Diffusion

·(edited)·· reads

AI TranslationSimplified ChineseEnglish

Key Insights

The content provides a guide on using Stable Diffusion and LoRA models for image processing. It begins by selecting a suitable photo with a clear subject and minimal detail. The process involves using a video tutorial to input prompt words into a tool called 'WD 1.4 Tagger' to generate keywords. These keywords are then adjusted and used in an image-to-image generation tool, where specific parameters such as sampler method, iteration steps, and redraw intensity are set. Four images are generated, and adjustments are made until a satisfactory result is achieved. The final image is then processed with upscaling algorithms, R-ESRGAN 4x+ Anime6B, at a scale of 4. The guide concludes by comparing the original and final images, showcasing the improvements made through this process.

Most of this post is based on this video

Intro

Resources used:

Stable Diffusion integrated package: https://www.bilibili.com/read/cv22159609/
Stable Diffusion model: https://tusi.cn/models/605039709506335625
LoRA model: https://tusi.cn/models/620507699483853069

How to do it

First, pick a suitable photo — clear subject, clear content, no fine details needed. Here's the photo I picked:
The photo I picked

Following the video, paste these prompts into img2img:

PROMPT

(white background:1.1),(simple background:1.1),(chibi),thick outline,happy,mugshot

Import the photo into WD 1.4 Tagger and you'll get these tags:
CodeBlock Loading...
Paste these tags into img2img, removing the ones you don't need.
Upload the image in the img2img tab and change the following parameters:
Parameter Value
Sampler DPM++ 3M SDE
Steps 30
Denoising strength 0.7
Batch count 4
Click Generate to create 4 images. You can adjust the denoising strength based on the results — a value between 0.6-0.7 usually works — until you get an image you like.
Send the image you like and its generation parameters to the Extras tab, set both Upscaler 1 and Upscaler 2 to R-ESRGAN 4x+ Anime6B, set Scale by to 4, and click Generate.
In the end, I got this image:
The final generated image

Parameter	Value
Sampler	DPM++ 3M SDE
Steps	30
Denoising strength	0.7
Batch count	4

Let's compare:

Most of this post is based on this video

Intro

Resources used:

Stable Diffusion integrated package: https://www.bilibili.com/read/cv22159609/
Stable Diffusion model: https://tusi.cn/models/605039709506335625
LoRA model: https://tusi.cn/models/620507699483853069

How to do it

First, pick a suitable photo — clear subject, clear content, no fine details needed. Here's the photo I picked:
The photo I picked

Following the video, paste these prompts into img2img:

PROMPT

(white background:1.1),(simple background:1.1),(chibi),thick outline,happy,mugshot

Import the photo into WD 1.4 Tagger and you'll get these tags:

PROMPT

outdoors, ocean, solo, day, black hair, 1boy, from behind, male focus, horizon, water, shorts, sky, photo background, rock, scenery, holding, short hair, standing, facing away, blue sky, pants, white shorts, long sleeves, shirt, shoes, bag, backpack, wide shot

CodeBlock Loading...

Paste these tags into img2img, removing the ones you don't need.
Upload the image in the img2img tab and change the following parameters:
Parameter Value
Sampler DPM++ 3M SDE
Steps 30
Denoising strength 0.7
Batch count 4
Click Generate to create 4 images. You can adjust the denoising strength based on the results — a value between 0.6-0.7 usually works — until you get an image you like.
Send the image you like and its generation parameters to the Extras tab, set both Upscaler 1 and Upscaler 2 to R-ESRGAN 4x+ Anime6B, set Scale by to 4, and click Generate.
In the end, I got this image:
The final generated image

Parameter	Value
Sampler	DPM++ 3M SDE
Steps	30
Denoising strength	0.7
Batch count	4

Let's compare: