In order not to produce pages with LoRA's trained on the same datasets with different Wan versions, I decided to put their descriptions in a separate articles.

Hollow knight [Wan 2.2 TI2V-5B]

https://civitai.com/models/1836374/hollow-knight-wan-22

Trigger Word: hollow_knight
Model: Wan 2.2 ti2v 5B
All examples are generated with 1.0 LoRA strength, CFG=5 and steps=30
For inference used Kijai's wanvideo_2_2_5B_I2V workflow with a small modification for t2v mode

⚠️ Work correctly only with resolution 1280×704 or 704×1280! ⚠️

Workflow Details

To load lora, add the "WanVideo Lora Select" node and connect it to the "WanVideo Model Loader". For t2v mode, switch the "Load image", "Resize Image v2", and "WanVideo Encoder" nodes to bypass.

Dataset Details

The training dataset consists of 18 videos with a length of 81 frames and 24 fps. The video are taken from this YouTube video with amazing music from the Hollow Knight game.

I didn't take "Pale Court" into the training dataset because it's too dark. The descriptions are generated using the Qwen web version.

Training Details

Initially, I wanted to continue eating cactus train the 1.3B version for wan 2.1, but due to the release of wan 2.2 with its new ti2v-5B version, my plans have changed. As soon as the diffusion pipe was updated for wan 2.2, I tried to train lora 5B on multiple resolutions, as I had previously done for Beauty of rain [Wan 1.3B] and Genshin TCG Style [Wan 1.3B]. But I got a broken lora that gave out creepy artifacts. I was forced to fix the video resolution in the training config.

ar_buckets = [[1280, 704]]
frame_buckets = [ 1, 81]

Only then did I get the correct results for the 5B version.