1980s (and some early 90's) style themes and photography Trained on a Data set of 136 images, V1 is more subtle (Lite), trained with just 3500 steps, V2 is stronger, bolder and more photo realistic, but less flexible, trained on 5000 steps, (6k steps attempted, but the results were starting to suffer from over training, so 5k steps was the sweet spot for V2)
The data set is very people centric, and will respond well to 80s hair styles and clothing, and some settings etc, You can prompt it with the following Sub styles,
'1980s, professional clean photography poster style,'
'1980s, clean photography style,'
'1980s, Garish Magazine Print style,'
'1980s, Faded vintage photography style,'
'1980s, vintage photography style,'
and then if desired:
'American' or 'British' (Some 'Japanese' too but it limited) - for different cultural looks,
Also tagged in captions: Curly hair, Perm hair, Bowl hair, 80's wavy hair, 80s Punk, 80s Glamour, 80s fashion, 80s Porn, 80s glasses, Aviators, hoop earnings, Shoulder pads, 80s high waist pants/shorts, 80s Jeans, Tan lines, Disco, 80s, office, 80s car /station wagon with wood panelling, 80s train interior, 80s living room, 80s couch, rotary phone, vintage television,
(contains some NSFW in the form of some mixed soft and hard 80s Porn (Amateur and Proffessional), but you wont get it with out asking)
Description
Use V2 for a stronger 80s aesthetic LoRa with higher Photo realism, And V1 (3500 steps)as a lighter softer version with better flexibility.
- I trained the same Data set in V1 to 6000 steps, but that seemed over trained in slightly degraded results. So for V2 here we have the 5000 step save point which seemed to be the sweet spot i my tests.
FAQ
Comments (13)
I love this, any plans for other decades like the 90s,2000s, or 70s, 60s, and 50s?
Thanks I was thinking about it yes, though I'm not sure which? Normally someone else makes these, but I just got impatient for a Krea 2 one to arrive, so decided to have a stab.
@franku196815 I think the guy who makes the vision vanguard loras is working on some. But one thing krea just cant do is the decade aesthetics.
@TheEventHorizon Did struggle to get the failed Polaroid look, I blamed that on it just getting drowned out by the cleaner images, as it was mix in my Lora, didnt realise it was krea2 Issue too. Only other persistent Problem i've noticed with Krea-2 is unwanted / unprompted patterns being placed everywhere, especially on fabrics, blank walls and upholstery etc. even if you ask for plain fabric or blank wall, it just ignores the prompt and does it anyway....
@franku196815 The worst is if you want make something sweaty and it gives you a damn waterfall.
@TheEventHorizon Yeah I noticed that when having someone on sunbed in front of a swimming pool, it insisted on making them look wet.
Have you actually tried era's with the model? It actually understands a lot of stuff already and gets the 'visual' vibe, you should try some tests with a scene and switch the year/decade in the prompt out, you might be surprised.. More LORA's welcome but KREA2 seems a lot better equipped than previous models in this regard
@MrTitsworth I have, but it really doesnt seem to get big 80s hair for instance, maybe im not using the right prompt, but even with enhanced prompt on it can never get the grainy look or fashion. It always looks like a modern photo.
@MrTitsworth Yes I have, and you're not, wrong - its not bad.
where can i find the 'krea2TurboUncensored' model in your workflow
I actually couldn’t even remember and had to enlist Gemini to hunt it down lol, there you go pal, https://huggingface.co/cusiman/Krea2_FP8/tree/main
Really nice - i also would like to see more decades. 70s, 60s, 50s, 90s
Yeah, Ill see what i can do, will have to spend some time building the datasets