Here’s a Himiko Toga Hunyuan Video LoRA I trained using 26 images at 512x768 resolution. Consider this a test lora. The whole process took about an hour. If you want to try making your own, check out: comfyonline one of the first platforms I’ve found that makes it really easy to train LoRAs for Hunyuan Video. Just go to explore and find the "Hunyuan Video lora train".

This Lora can produce both realistic style and animated style as well as a mix of both, (needs more training on realistic, might struggle).

It was trained on SFW and NSFW images. I recommend adding: "T0gA_H!m1k0, anime style, blonde hair, bangs, twin buns, yellow eyes, sharp teeth" or "T0gA_H!m1k0, realistic style, blonde hair, bangs, twin buns, yellow eyes, sharp teeth" At the start of your prompts, also please feel free to check each example video as I will leave the prompts that created them.

I recommend using the following settings, I'll also leave my workflow, (its the dataset download, wasn't sure how to upload the workflow json).

Base model: hunyuan_video_720_cfgdistill_fp8_e4m3fn

Lora strength: 1.00

Steps: 30

Resolution: 512x768 works well experiment with it.

embdedded_guidance_scale: 9 works well experiment with it.

flow_shift: 12 works well experiment with it.

num_frames: 129 works well experiment with it.

FPS: 24 works well experiment with it.

denoise_strength: 1.00

Recommend using: Enhance-A-Video weight 7.00, turn single blocks false and double blocks true

Sageattention: enabled.

Make sure you add "hunyuanvideo Lora block edit" node and DISABLE all single blocks, keep only the double blocks, connect your load lora node to this node.

Custom prompt template node added with default system prompt.

hunyuanvideo cfg node (negative prompt) added with: "still frame, no movement, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion".

I hope you enjoy it!