Janine LoRA

Character LoRA of Janine, an AI-generated (synthetic) person. Trained on a 40-image dataset with detailed natural-language captions. This repository contains adapters for two different base models — pick the one that matches your generation pipeline.

Trigger word: janine — include it in every prompt, e.g. janine, a woman walking in a garden.

Files

File	Base model	Use for
janine_flux_lora.safetensors	FLUX.1-dev	Still images
janine_wan22_high_noise.safetensors	Wan 2.2	Video (high-noise expert)
janine_wan22_low_noise.safetensors	Wan 2.2	Video (low-noise expert)

A LoRA only works with the base model it was trained on — the Flux file will not load on Wan, and vice versa.

Usage in ComfyUI

Images (Flux)

1. Put janine_flux_lora.safetensors in ComfyUI/models/loras/.
2. Load a FLUX.1-dev checkpoint, add a Load LoRA node.
3. Suggested LoRA strength: 0.8–1.0.
4. Prompt with the trigger, e.g. janine, a woman wearing a white shirt, soft window light, close-up portrait.

Video (Wan 2.2)

Wan 2.2 uses a MoE (high-noise + low-noise) architecture, so this repo ships two files.

1. Put both janine_wan22_*.safetensors in ComfyUI/models/loras/.
2. In your Wan 2.2 workflow, load the high_noise LoRA on the high-noise model and the low_noise LoRA on the low-noise model.
3. Prompt with the trigger janine.

Tip: for image-to-video you can also generate a still with the Flux LoRA first, then animate it with a standard Wan i2v workflow — no video LoRA required.

Training

• Dataset: 40 images, captioned in natural language, each starting with the trigger janine.
• Flux: ~2000 steps, learning rate 1e-4, LoRA rank 16.
• Wan 2.2: ~2000 steps, learning rate 2e-4, LoRA rank 32.
• Trained on WaveSpeedAI.

Notes

Janine is a fully synthetic, AI-generated character and does not depict a real person.