Description

I am very happy to share my magnum opus LoRA, which I've been working on for the past month since Wan came out. This is the best LoRA ~~on Civitai~~ I have ever trained, and I have to say once again - WanVideo is an amazing model.

I'm currently writing a detailed post about the training process and will update the model page soon enough.

I'll also be adding TONS of showcase videos this week and next because, honestly, I keep getting high-quality clips from nearly any prompt I feed into it.

I'm sorry in advance for the number of visually similar clips in the gallery - it's just that I usually generate three clips per prompt, and most of the time, all 3 turn out perfect (from my POV, of course). I just can't decide which one is the best, so I end up keeping them all.

This LoRA was trained for ~90 hours on an RTX 3090 with musubi-tuner using a mixed dataset of 240 clips and 120 images. It could have been done faster, but I was obsessed with pushing the limits to create a state-of-the-art style model. It’s up to you to decide if I succeeded.

Usage

The trigger phrase is Studio Ghibli style - all training captions were prefixed with these words.

All clips I publish in gallery are raw model outputs using a single LoRA, without post-processing, upscaling, or interpolation.

Workflows are embedded with each clip. You can download example workflow (JSON) here: https://files.catbox.moe/1nrkms.json

I apply a lot of optimizations, including fp8_e5m2 checkpoints + torch.compile, TeaCache, Enhance-A-video, Fp16_fast, SLG, and (sometimes) Zero-Star. Rendering a 640x480x81 clip takes about 5 minutes (RTX 3090).

Dataset

🚧 (Coming soon...)

Training

🚧 (More details coming soon...)

Studio Ghibli Style 🎥 Wan2.1-T2V-14B - 1.0