~ Version 1.0 ~
Prompt for WAN 2.1: A man with glasses. He is wearing a blue shirt. Anime style.
Prompt for WAN 2.2: A fat man with glasses. He is wearing a blue collared shirt. Anime style.
Supported models:
- WAN 2.1 T2V 14B fp16
- WAN 2.2 I2V 14B fp16
- WAN 2.2 T2V 14B fp16
I included workflow examples for all the above models + explanations on parameters in a separate article: https://civarchive.com/articles/18711
Recommended samplers:
- Wan 2.1 UniPC Scheduler: Simple Lora strength: 1.0
- Wan 2.2 Euler Scheduler: Beta Lora strength: 1.0
Prompt tips (green - positive, red - negative):
- Anime style - required to keep the style, you can make him fully realistic if you remove this tag and use Realistic instead
- A fat man - depending on your cfg/step count/resolution/prompts he can sometimes appear more skinny than intended. Including "fat" or "chubby" will help retain his body type
Most of the I2V source images can be found on my Seiichi Kinoshita Illustrious page.
~ Version betaWan ~ (old)
Main prompt: A fat man. He is wearing a blue collared shirt and glasses. Anime style. <------ should always be included
Negative prompts:
色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走Supported models:
- WAN 2.1 T2V 14B fp16
- WAN 2.2 I2V 14B fp16
- WAN 2.2 T2V 14B fp16
(The Lora should work with the fp8 versions as well but I haven't tested it.)
Recommended samplers:
- Wan 2.1 UniPC Scheduler: Simple Lora strength: 1.0
- Wan 2.2 Euler Scheduler: Beta Lora strength: 1.0
Note: CFG/Steps will depend on what kind of workflows, models and loras you use (e.g. Lightx2v). Refer to gallery videos comments for examples and parameter recommendation.
Prompt tips (green - positive, red - negative):
- Anime style - required to keep the style, you can make him fully realistic if you remove this tag and use Realistic instead
- He remains silent as his mouth is closed - prompts like this help the constant 24/7 talking
Most of the I2V source images can be found on my Seiichi Kinoshita Illustrious page.
Description
This is my 1.0 release of the WAN 2.1 T2V 14B Lora for Seiichi Kinoshita
Changes in 1.0:
- trained using Musubi Tuner instead of AI-Toolkit
- bigger dataset with higher resolution training (results in better details for full body/wide shots)
- better prompt adherence
- better motion
- less talking/can be made completely silent with help of WanVideoNAG node
I decided to do another WAN 2.1 version because training a single Lora is faster compared to WAN 2.2's high/low model requirements. Though doing a WAN 2.2 version is my next goal.
The Lora is not perfect, it can be further improved by increasing resolution training/fine-tune captions/adjusting dataset but I will save that for the WAN 2.2 version.
Like my previous Lora this will work with WAN 2.1 AND WAN 2.2, however, WAN 2.2 compatibility isn't 100%. By default the model will make Seiichi skinny which requires "A fat man" to be always included in the prompt.
This time I am also attaching ComfyUI workflows that can be use right away with the Lora.