Wan V1.1:

Objectively, this might look worse than v1.0. But it's closer to the style I wanted to go for, and less "Wan" default look. It's using the exact same dataset, but training regime has changed.

Does not work well with lightning loras, or I didn't find good settings. Tends to lose style and become oversaturated.

Original:

The lora I was afraid to make. Not because I didn't want it, but I was afraid of failing. So much of my childhood is contained in this lora.

It's trained on images < 400p and will lose strength the higher you push the resolution. I've mostly tried conditioning between 8 and 10.

You don't really need the trigger word, imo, but it could be used to channel some more style into the generation

Does not do so well with combat scenes. Body horror is likely to ensue. For an update I will add some appropriate clips and see if that helps.

For prompt hints, check the example generations I've posted. Here's also one from the training data, to give you a sense of the format:

"A group of robed figures, likely a council of wizards or sorcerers, stand around a glowing, pink crystal at the center of a circular stone table. The crystal emits a mystical aura, and the figures appear to be discussing or casting a spell. The background features a dark, cavernous interior with stone walls and alcoves containing various objects, including a large, skeletal creature. The setting suggests a magical or ancient ruin."

All showcase images/videos contain ComfyUI workflows.

Wan 2.2:

The Low noise lora has absolutely most effect for style. High noise still affects the outcome, but to a much less degree.

For the showcase, I use 25 steps: 15/10. 0.8 strength, and dpm++_sde scheduler in ComfyUI. I haven't had a lot of time for testing, but it might be that it loses strength as you go up in frames.

I use Q4.gguf for High noise, and Q6 for Low noise, for no particular reason. No block swapping.

Wan lora trained with diffusion-pipe.