Video Age Slider
This LoRA is basically a little time machine for your video.
The strength controls how "old" the footage looks. Go negative and the video gets cleaner, sharper, crispier and more digital, with stronger contrast and a modern high-quality look.
With positive values, things become softer and more natural, with less of that ultra-clean digital look.
Push the strength into positive territory and things start going downhill, in the best possible way. The image gradually gets softer, blurrier and more analog, with faded colors, lower contrast, video noise and other characteristics of older recordings.
Keep pushing it and eventually you're deep into VHS territory: washed-out colors, blurry details, chromatic bleeding, tape noise, scan lines, tracking problems and that wonderful "this was recorded in 1987 and copied 14 times" look. 📼
Roughly:
Negative values → crisp / modern / high contrast
Low values → softer / natural / realistic
Positive values → old / analog / VHS
High positive values → very old / very blurry / VHS went to war
The exact sweet spot depends on the model, prompt and video, so don't treat the values as strict presets. The fun part is sliding through the range and seeing where the video starts looking like it was recorded by a professional camera... and where it starts looking like someone's dad found an old camcorder in the garage.
Description
First version
FAQ
Comments (21)
That's so cool, thank you ❤️
Thank you for your kind words.
This is amazing. I've been working on a dataset for old tokusatsu shows, which were filmed on film and had hand-animated optical effects. I'm thinking a slider lora like what you do would be able to target it better. Same goes for a cel animation lora to cover legit old animation techniques (like not having CG backgrounds and using parallax instead, etc...), against modern techniques.
Anyway, this lora result makes me more excited to try to train this. I wrote a script for musubi-turner based on your reddit post that I hope will work. Claude claims that it found ways to improve upon what you posted, but who knows if that actually works or makes it worse.
Hey I hope your tokusatsu lora goes well, I'm very interested in that and things like Space Sheriff Shaider etc style.
Lots of Japanese genres that would be cool really e.g things like Sukeban shows like Sukeban Deka, Hana no Asuka-gumi! etc have cool styles too. (Not telling you to do anything haha, just mentioning how many untapped styles there is out there other than typical hollywood)
@Light6969 Hahaha most of my dataset is Space Sheriff Shaider specifically. I'm happy to hear that's what you want to see. I don't know if you've seen my account, but I'm trying to pull together something that can achieve convincing shots from a non-existent 80s sukeban henshin show, with monsters. I want something more classic sukeban than stuff like Sukeban Deka or Asuka-Gumi, though I love those. I definitely want some "Girl in the iron mask" vibes from Sukeban Deka 2, but again, with more convincing, ACTUALLY delinquent gang sukeban.
I used to make a series of this on Sora 2 when it was active, even getting into a documentary on a pivotal scene in the show, and an anime adaptation. Definitely more manual work with Minimax, but that's actually exciting to me, that I'm forced to do more creative work with it. Makes it feel more like my own.
So yeah, thanks again for the info you've shared on your training process. I hope my copy of it works out. I think what you're doing is underappreciated, considering how smart it is to make it concentrate only on the exact changes you want to make, minimizing the training of unintended concepts, since you'd preferably make mirrored datasets, and it only targets the differences. It's a really smart way to laser focus a concept. I just hope it works out for me.
@Jellai oh wow really!? I've been watching it lately, I love the girl in this show and the crazy enemies.
Yeah I looked at your profile after that comment and saw some obvious Sukeban x tokusatsu inspired videos, really cool! haha
I get exactly what you mean with Minimax, the model makes you really start to think like an actual director, which is really fun (and addictive) and satisfying.
Oh by the way I'm not the training or lora creator guy haha, just a passerby, I'm curious though, which reddit thread was it? ?Sounds like good info
@Light6969 Oh wow. Your comment exposed just how deeply confused I am. I'm so busy at work, and I just jumped in and info dumped without actually paying close attention to whose lora this was, or who was talking with me.
So yeah, this is the thread I'm talking about:
https://www.reddit.com/r/StableDiffusion/comments/1ku2ijc/discovery_creating_sliding_loras_for_wan_and/
@ComfyTinker was the one who wrote it. It's just an overview of their approach. I basically vibecoded a script for musubi-tuner that did what was described there, with some enhancements that it said would make it work better. I don't know if @alcaitiff uses this method. Sorry, both of you, for mixing you up. Anyway, I guess we'll see if Claude actually understood musubi-tuner and the principles in that reddit post enough to do something that actually works.
The more I looked into it, the more I think it's brilliant. The big down side is that it requires you to train TWO loras, but the strength of that is that it laser targets the EXACT concept you want, and everything else gets tossed out as redundant. Then when you apply it to Minimax, it's more likely to leave all other concepts untouched. So you could make it so that it doesn't result in different faces, or higher color saturation, or any number of things that are often accidentally done by loras.
@Light6969 oh, also I found I was already following you. I'm a big fan of your work. I see the interest in old media, like me.
@Jellai lol yeah I had a feeling it was some confusion like that, no worries I do similar things a lot lol. Thanks for the reddit thread, I see it's an older really cool idea that you are trying to adjust and improve, hope it goes well!
Lol thanks, yeah we have similar taste it seems, oh yeah I should follow you too!
OOohhh.... FOOTAGE age... Yeah I was going to say this didn't work at all even at -3.
Its not making video futuristic 32k at -3? Meh, useless :p
Thanks, love the idea. I'm curious to try an old 80s reference image and then going -3 to see if you get the same scene but more modern looking lol
It worked, it created a nice look, where the colours were high contrast but with a modern look.
@Light6969 Please publish so we can see it =]
@alcaitiff I may have jumped to conclusions, it's more like the same character reference but when I prompt for a different scene basically rather than the exact same scene, which works well. I'll show an example of that though!
Thanks For H3
Very interested to see if this can clean up older footage, and how well it does that.
It's crazy. You can use old blurry and super dark videos, use reference pics of the character seen and prompt it to be a good quality and bright video and you get like the same video, just in good quality. This works without this lora, but at -5 it adds even more bright and realistic visuals.
This is the first time I've ever commented on a LoRA, I just wanted to share my appreciation for your work on this, it's quite excellent. For anyone trying it, take a look at the prompts in the examples - the LoRA is very responsive to detailed descriptions of the aesthetic, including specular highlights, camera jitter, lens effects etc.
Thank you so much for your kind words, @milowa505. They truly mean a lot to me!
This is amazing thanks!