Version 2 is not perfect either and for some prompts version one can still be better. Version 2 may give results sometimes where they look childish. Still testing why and how. Don't usually get that if I use miniature man or miniature woman.
Don't forget to follow to get notifications of updates.
Warning: This LoRA is not perfect, I do plan at least one more version, but after many hours of work and many early versions for testing and creating better datasets I have decided to publish it.
I would love to see anything created so please feel free to post your creations here. Will help me improve on the future versions.
This LoRA is designed to overcome the problem of plastic looking figures when trying to get miniature people. All of the models I have tried over the last 1 1/2 years since I first tried getting images like these I could never get what I wanted. I will warn that this is my first version and is not perfect by far. It can have issues with certain prompts and does require prompting that describes the scene like a miniature person.
For example if you just prompt for "miniature person standing next to a truck" you might get a smaller version of a person but unlikely to get a really small person. But "miniature person standing next to a truck tire" will get a much smaller person. Also mentioning them being on a kitchen counter will most likely get you a miniature person just saying a miniature person next to a counter may not. You can use actual celebrity names for example as well but sometimes does not do as well.
Does not do closeups well so far in my testing. If you do not describe clothing you are very likely to get plaid. It was worse in earlier versions but that still happens.
I will add more info as I learn more. But will suggest checking out some of my prompts used in the example images for ideas.
Considering the amount of work I put into making this a reality if you would like to buy me a cup of electricity it would really help. https://ko-fi.com/kleebztech
If interested in how I created this I did create a video:
I did train it with Flux Gym which I have a tutorial on using.
Description
FAQ
Comments (16)
This looks fantastically well done.
I advise testing it on the on-site generator ASAP though, as I've seen a few Flux Gym trained LoRAs fail there over the weekend. Hopefully it's fixed/not affecting you, but best to know asap!
Editing to confirm it worked for me in a trial!
I did give it a quick try and seems to work. Thanks!
Thank you so much for making this! I've been (slowly) tinkering with trying to train a similar lora and this helps SO much to refine further and even get a better understanding of how flux processes scale.
Looking forward to watching the video you mentioned!
Welcome! Yeah I am still in search of getting the scale consistent. But wanted to at least get something that worked and hope to improve it. Hope to get the video done soon but will warn, I don't really have any answers on the best way of doing things yet and I may even being doing it the wrong way. I do plan on experimenting more soon with more versions once I can build up a better dataset.
@toasters No worries or rush at all about what you share or when--it's already super helpful AND good "knowledge" even if we're doing it "wrong" haha! Flux is an odd duck (but so am I)
I downloaded this beautiful model the day it came out, and today I watched your YouTube tutorial on how you made it. Beautiful work! Thank you so much!
Thank you!
I have been trying to figure out to get mini people for quite awhile. Your Lora did the trick in 5 tries. Pretty cool. Thanks!
Thanks! Yeah I have been myself for some time. It takes a little bit of learning to prompt but once you figure it out it does decent with many things.
This is adorable model! 🌸
It truly sparks a lot of inspiration. ✨
I was not interested in Realistic models before but when I saw your model I fell in love with Realistic works.
Thank you.
Very good model. It does the job more than half of the time. On other occasions, the 'miniature person' may have a duplicated head or a dismembered body, so perhaps future versions (more training?) will lower this probability. But it's already a remarkably useful lora in this first iteration!
Thanks for the input. I have found it often depends on the prompt. Some do well and others not so much. Would be curious which version of the Flux model you are using. I am working on the next version now.
@toasters I use the Flux Dev from CivitAI (https://civitai.com/models/618692/flux) while I'm away from home, as is the case now. It's based on this checkpoint that I made my comment. Otherwise I have a Comfy setup, but I haven't tested your Lora on that yet. Best of luck with the next version!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.