WAN 2.2 Fantasy Theme LoRA
A fantasy-themed LoRA trained on artwork featuring elves, witches, and mystical creatures, with light NSFW elements.
Trigger Wordfantasylora
About
This LoRA focuses on a rich fantasy aesthetic with glowing environments, ethereal lighting, and mythical tones. Perfect for fantasy character portraits, enchanted landscapes, and magical storytelling visuals.
Usage Tips
If image quality degrades, adjust LoRA strength for the low noise model.
Increase high noise strength for stronger style alignment.
Decrease low noise strength to improve detail.
Experiment with different samplers and schedulers for unique looks.
Recommended Settings (ComfyUI)
High Noise Model
Sampler: LCM
Scheduler: Beta
Steps: 0–14
CFG: 3.5
Low Noise Model
Sampler: Euler
Scheduler: Beta
Steps: 14–24
CFG: 3.5
Example Prompt
“fantasylora, a young woman with silver hair that shifts between silver and deep sapphire, her skin like polished river stones, standing at the edge of a river glowing with liquid starlight. She holds a staff carved from a single moonbeam, its tip pulsing with blue light as shimmering bridges form over the water.”
Community Contribution
I’m working on future versions of this LoRA and would love your help improving it! If you create great renders using WAN 2.2, please upload them to help build a better training dataset.
→ Submit your images here
If you’d like to support future work and new LoRAs, you can:
Use my RunPod Cloud GPU referral link. RunPod offers a wide range of cloud compute options for all kinds of open-source AI projects. When you sign up using my link and deposit $10, we’ll both receive an extra $5 in RunPod credits. Every bit of support helps a lot with training these LoRAs!
https://get.runpod.io/3cbr613qjf6gJoin my Patreon. I share a variety of installation guides and one-click installers to simplify setting up local AI projects—including image and video generation, text-to-speech tools, and more.
https://www.patreon.com/cw/TheLocalLab
Enjoy the LoRA and share your creations!
Description
WAN 2.2 Realistic Fantasy v1.0 High Noise LoRA Model Version
FAQ
Comments (6)
Interesting... I hadn't considered changing the Sampler and Scheduler to be different between the High Noise pass and the Low Noise pass... I need a second computer.
the High model sets the video, like a placeholder. adds shapes and movement,
the low model is essentially a refiner, adds details to the noise generated by the high model.
you can also:
pair Wan 2.2 original High with a boosted speed Model from here/Huggingface
the opposite works too
pair A GGUF High with a safetensor low
the opposite works too
depending on your requirements - only apply high loras - to set shapes or lighting, or like a character, or do only low loras to add detail to the high models generated noise, instead of the pair, using both offcourse keeps the generation to what the lora was meant for if theres 2. I noticed I prefer using low loras more.
what I found is that, original wan 2.2 with lightx2v lora (doesnt matter which one, as long as its for 4 steps, works better if its paired with a merged checkpoint from civit. depending on the look you want - I treat it like these checkpoints are for images, each have a different aesthetic. since original wan2.2 high has good knowledge of how to set the videos shapes lighting and movement, using low+loras that add details as a refiner for details, videos come out really nice.
and even though Im using the lightx2v lora for high, and a speed checkpoint for low, 2+2 steps work, but to get a really clear 480x832, I split 30%/70% steps high-low with around 4 or 5 high and 12 to 14 steps low - defeats the purpose for using the speed loras, but extra time = quality. 720 takes too long, so I stick with 2+2 or 4+4
So everyone knows, this lora is far better than the example images. Check out the post on reddit. Someone made a new set of photos with the same prompt and it looks a lot better. There's some good training here.
Yes and this is only the first version. I'm shocked no one has posted any new images despite there being many downloads. I would love some better images to make another lora. The original dataset wasn't really the best to begin with and belonged to someone else but still helped produce decent images.
I suspect the better quality results used the res_2s sampler and bong_tangent scheduler.
@TheLocalLab In my defense, I reflexively download, and give a like to, anything I come across that piques my interest because I've gotten burned too many times in the past when someone rage quits and takes all their models with them. As a result, I have a massive backlog of things to play around with. But, I am limited by what little free time I have and my budget hardware, especially when it comes to videos. Also, despite starting out with SD1.4 and being on Civitai site since the early days (under a different account that I'm not ashamed of to share with my family) I still don't consider my stuff to be very good quality so I tend to lean into funny instead.





