Jenn (Reference Lora) Hunyuan video LoRA

Jenn (j3nn) is a fictional person created to research Hunyuan video LoRA training.

She is a 30 year old woman. She has fair skin, dark hair with auburn highlights, and is of European and East-Asian descent.

The goal is to produce a high quality LoRA trained on synthetic images (ideal conditions) to use as a reference when training a LoRA that utilizes real images and video (less than ideal conditions).

A properly trained character LoRA will:

• consistently produce a likeness of the character

• preserve the original video as much as possible

• not introduce glitching

• follow prompts without biasing the video to the training data

• be easy to use and not require a specific prompt or extreme guidance value

• play nicely with other properly trained LoRAs

• respond linearly to strengths between 0.8 and 1.2

Jenn was trained on 35 images with ~1.6MP resolution (1280x1280, 1024x1536), at a resolution of 1MP ([1024]) at a learning rate of 1e-5 for 1760 steps with a gradient accumulation value of 4.

--

This model is useful for finding ideal settings when combining multiple LoRAs. Dialing in the correct strength values of each LoRA is often tedious, and it can be made more difficult when both of the LoRAs are over-trained.

You can rest easy knowing this LoRA is correctly trained. If you want to produce a LoRA that combines nicely with character LoRAs you can use Jenn as a reference. You can utilize Jenn in your example images because she's not real.

I would love feedback.

Note: I retain commercial rights to Jenn.