Updated to Wan 2.2
Rapunzel from Tangled
"Rapunzel" is the trigger.
Very important! Start your prompt with "3D Cartoon Style", unless using another style lora
Movie outfit
She is wearing a purple Dress with puffy sleeves, barefootMovie Hairstyles
Very long flowing golden blonde hairVery long golden blonde hair braided with colorful flowersShort tousled brown hair"Green chameleon" for Pascal
Use both Low and High noise loras.
All preview videos made with Wan 2.2 lightx2v lora on Q8. Comfyui workflows embedded in videos
Expect poor results with 2.1 lightx2v. I use the official Wan 2.2 lightx2v lora found here: https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main/Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1
Not tested extensively with other 2.2 loras but should work okay, adjust strength as necessary
Description
FAQ
Comments (9)
I've been gooning with movie and TV show accurate Rapunzel porn for 15 years now, spending like 5k on her commissions alone...
But I think we have reached peak TangledPorn. I don't think we can go much higher than this. Great work dude, you are the goat.
I am really happy for you
I'm having trouble getting likeness with anything other than the basic example prompts demoed. Do you mind posting one of your successful gens with metadata so I can see your workflow?
@Clocksmith Hey, sorry you're having trouble. It is crucial to start your prompt with "3D Cartoon Style" and use both high noise and low noise loras
Long prompts with lots of detail should not affect likeness, I've posted a few example videos using more complex prompts with workflows embedded, hopefully they help.
https://civitai.com/posts/22519183
Motion adherence isn't perfect but this is down to lightx2v.
My workflows are quite simple, native comfyui nodes and Rapunzel+2.2lightx2v loras, nothing special.
@Kong__ Wow. Those are some fantastically complex and beautiful prompts. Thanks for taking the time to put those together.
So I was definitely using "3D Cartoon Style" and both loras. I know I had that all correct because I was able to duplicate your stock examples with just the prompts.
What wasn't working for me was a very simple NSFW prompt of "3D Cartoon Style. Rapunzel is posing seductively and then turns around and looks back. She is completely naked." The likeness almost completely dropped and it looked almost like a knock-off generic 3D blonde created by a third-party studio trying to imitate Rapunzel without infringing on copyright.
The results are even worse if I try to combine with a NSFW lora like playtime_ai's cowgirl lora. It just ends up inserting a generic blonde. I can get some likeness if I only use the high cowgirl lora with strength at 0.5, but then movement is completely diminished.
@Clocksmith That's really weird, I've not encountered any prompts that give issues with likeness. I've tested your prompt (slightly edited to put more emphasis on nudity) several times now on Q8 and fp8_scaled, lightx2v on and off and had no problems.
Low shift values could potentially introduce some likeness issues, I recommend 5.0-8.0. If that still doesn't help try being more descriptive, include her long golden blonde hair and green eyes in your prompt.
As for combining with other loras it can be tricky you need to find the right balance of lora strengths and including the trigger words from other loras. Examples of your prompt and Rapunzel lora combined with a playtime_ai lora here
https://civitai.com/posts/22530504
playtime_ai lora was set to 0.9 on high noise and 0.7 on low noise
@Kong__ Thanks for the continued debugging and excellent examples.
My workflow uses shift values at 8 and I have never modified them.
I don't use quantized models but I doubt that makes a difference unless you trained her on a quantized model?
I dug my way to your mp4 for the hand job. Strangely civit had me downloading it as a "noise" file when I tried downloading it directly. I have your workflow loaded but you are referencing loras called "high_noise_model" and "low_noise_model". Can you tell me what those actually are?
@Clocksmith Trained on fp16, I've tested the lora on various quants and fp8/fp8_scaled both native comfy nodes and KJ nodes and confirmed it works on all, it really shouldn't require anything special to use.
high_noise and low_noise are just the lightx2v loras I use, I should probably rename them to avoid confusion. They're found here
But any 2.2 lightx2v/lightning lora should work
This user has had success
https://civitai.com/posts/22551298
Perhaps you could try their workflow? If that fails I'm all out of ideas, maybe you could share your workflow?
Impressive! The Bar to work with this stuff locally is so high and the cost of GPU has gone up 80% instead of down.. Wish we could get more hype but this stuff def has a future.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.