
LoRa of Canadian Politician Mark Carney. Include "Mark Carney" in the prompt, strength 0.8 to 1.0.
Mark Carney is a formidable man. A bone fide globalist, he has served all over the world: interned at Goldman Sachs, governed the Bank of England, is a member of the "Group of 30", helped coin the term "The Great Reset"...
He even did a stint at the Bank of Canada, his home country.
I wish him well in his bid to replace Justin Trudeau, and hope that this LoRa honors Mark as Mark has honored Canada.
PROCESS: This was pretty low effort flux lora, and yet it works, so I'm going to share my process:
1. Collect ~25 images
2. Upload to Civitai, run the prompt generator on default settings
3. Edit each prompt to reference the character's name.
4. Remove (most) descriptions of the character's clothing, hairstyle, etc.
5. Add a few regularization images. i.e. pictures of a similar style/category but not of your character. Tag them similar to your other images, but without the character name.
6. Halve the UNET learning rate, double the epochs to 10.
NOTES:
Steps 1, 2, and 3 are probably enough.
4 must help: On the one hand, you want your model to describe the things you don't care about, on the other hand, you want it to describe your photo & be similar to what you'll prompt when generating. The compromise is to have your character's name, but otherwise only have a short description of the scene. Civitai does this pretty well, and removing the character description makes their prompts more concise, which probably helps.
5: I've heard you don't need regularization images, but a few can't hurt. That said, my model has trouble depicting multiple characters anyway, so perhaps you either need ~25 of these, or just have to settle for a model that only portrays your character.
6: I've read low-image count flux loras are best served by having many training steps, counter-balanced by a low learning rate. More training steps lets your model learn the details (exactly what their face looks like), and lowering the learning rate proportionally counteracts the negative effects of overtraining.
If true, raising the epoch count seems to be the way to increase training steps on Civitai. But my 5-epoch model also works about as well as this 10-epoch model, so this step can probably be skipped if you don't want to fiddle with settings.
7 (bonus): Just check the "this is not a real person" box before generating. This is a public figure, civitai already has Donald Trump loras, because this is fair use. If Mark Carney's lawyers are reading this, I would love to start a high-profile legal case with you. DM me.
描述:
Works best at strength 1.0, probably fine at 0.8. Haven't run into problems.
Training prompts all included "The image depicts Mark Carney..." or similar, including that name somewhere should be all you need.
训练词语: Mark Carney
名称: 1289557_training_data.zip
大小 (KB): 2317
类型: Training Data
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
名称: carney.1.safetensors
大小 (KB): 18824
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success