This LoRA makes everything in the scene more, or less, luxurious. The LoRA affects everything in a scene except people but it will affect their clothes.
The number on the preview videos is the LoRA strength. The videos under the high model have random seeds and the videos under the low noise model all use the same seed.
This is a slider LoRA. It was trained with the WAN 2.2 model only, without any training images or videos. The LoRA simply enhance concepts already presents in the model.
You should be able to use it with other LoRAs without interference since it doesn't contain anything that's not in the base model already.
Maximum Strength: 7. You can push it higher but it will affect colors.You might be able to mitigate that with prompting or by lowering the high noise model strength.
Minimum Strength: -11. Lower than that degrades the video quality too much.
The positive training prompt is "luxurious opulent sumptuous expensive rich costly deluxe lush splendid magnificent lavish"
The negative training prompt is "poor impoverish humble cheap modest undistinguished meager worn"
Target class: (empty/everything)
Anchor class: person, video quality
(My first attempt degraded the video quality a lot with negative weight so I added "video quality" to the anchor and it worked well. If I retrain it, I will anchor sunlight and emotions as well to try to avoid some of the side effects).
If you are interested in making something similar, I created this with Ostris AI-Toolkit. https://github.com/ostris/ai-toolkit
Video tutorial by Ostris:
Description
Samples in the Low noise gallery all use the seed (42).
All houses have the same prompt except for the number on the sign.
All cars have the same prompt except for the number on the car.
All astronauts have the same prompt except for the number on the sign.
FAQ
Comments (31)
Judging from your previews, you have an interesting concept of what "luxury" is.
Technically not "his" idea, but the model's - sliders are trained on prompts only, whatever is created is how the model perceives the contrast between the positive and negative value prompts. But yes, it is quite fascinating.
That's right. The samples are the first ones I got with a legible number on the books. It's basically what Wan thinks luxurious or cheap is at various weight.
Nice idea
Great idea. Would it be possible to do a loose clothing slider?
Probably yes, it just need to be a concept Wan already knows about. I think that would work. Thanks for the idea.
I made one. I think it turned out pretty good: https://civitai.com/models/2015480?modelVersionId=2281140
@qcgooner Amazing!! Yes this is incredible thank you so much!!! Impressive speed, too.
can you please make one for dreamcore ? non has been done for dreamcore
I need two opposites concepts to train the LoRA. I'm not sure what the opposite of dreamcore would be and if Wan understands this already. Perhaps dream to realistic or dream to nightmare would work. Still a cool idea.
Were all the samples generated using the same seed?
Also, some samples include a negative prompt while others don't; this makes it difficult to accurately evaluate the effect of LoRA.
I had to generate them using different seeds because I couldn't get the number on the book when using the same one for all the samples. I will try to generate some more with a static one.
For the negatives prompt, they should be all the same. I'm not sure why some disappeared. I will update them.
I added some fixed seed videos to the low noise model gallery.
@qcgooner
This is a very good example. Thank you.
It would be interesting if there were also LoRA for things like the Skin Tone Slider, Depth of Field Slider, People Count Slider, and Hair Length Slider, which Mr. ostris created a while ago. https://civitai.com/user/ostris/models
So how do you train this kind of lora? Sounds interesting
You can't just say it's easy to make, with no training data, and then just not drop any info :D Share the deets, man!
The dataset is just 2 short prompts. I added them tot he model description and a link to the tool used.
This is cool, what a neat concept. Thanks for making this.
What I learned doing this, Wan thinks that:
Rich people are happy and inundated by sunlight.
Poor people are sad and moody and don't get sunlight.
Expensive cars are sport cars.
Cheap cars can't move.
Expensive astronaut is a regular astronaut.
Cheap astronaut is still on earth.
Some of that is not too far from reality!
i just keep thinking poor -11
Oh, but can it do lowcost cosplay?
@firemanbrakeneck It's never going to be as creative as that low cost cosplay dude but it kinda works. I added a few videos to the Gallery.
@qcgooner Ha, way to go.
And who knows, maybe someday.
Thx, for your work:) Yours and alcaitiffs Sliders are extremely useful. Is it possible to make sliders for saturation, light/dark, contrast, colour and so on?
I made one for overall scene brightness: https://civitai.com/models/2045509/brightness-slider-wan22
Thanks for the ideas. For color corrections there's a lot of possibility. I can't make one for every color pair but but I was thinking making one for cold to hot colors. I have a lot of generations that end up with a blue tint and I think a color temperature slider could help with that.
@qcgooner thx for your reply and for the brightness slider. I am not an English native speaker so I was not very exact. I was more thinking about a slider for saturated, strong, a lot of color and pale, not many color, nearly black and white. Its like your brightness slider but without changing brightness :)
I hope I don't annoy you but I had some more ideas for a slider. Maybe a weather slider? from heavy rain, stormy, fast moving clouds to sunny, cloudless, heatwaves in the air. Or a vintage/futuristic slider, or maybe for different ages like medieval/now, or a funny one, a animal,pet/monster,alien slider. :) I really like sliders, they makes prompting so much easier. I didnt look into it but is it possible to make the strength random?
wouldn’t it be nice if someone can make a random prompt generator with sliders?
@dft78750707 No worries. I really appreciate the suggestions. I think understood your first suggestion it's just that making the sliders is kinda like cooking and I'm not a chef. I just throw stuff in the pot and sometimes the model makes something unexpected and I lean into it. It's still good to have many ideas to try, I take them all in note.
Some concepts are also easier to train than others. For example, I know many people want a skin tone slider but I just can't figure it how to make one that works properly.
For making the strength random, I never tried it but I'm sure there is a few custom nodes for that in ComfyUI. This one for example: https://github.com/unwdef/unwdef-nodes-comfyui
@qcgooner For making the strength random I already found a solution and its quite easy, just use the "Random Number" node from was-node-suite and add that to the strength connect point of the "WanVideo Lora Select" node.