I used prompt-engineering techniques with Flux2 to generate a synthetic dataset of images that replicate the style found in certain online communities—without using real people’s photos. This ensures the dataset is ethically sourced and avoids any issues related to identifiable individuals. And it worked strikingly well.
It pairs well with ComfyUI-Impact-Pack’s Iterative Upscale nodes to gradually enhance resolution and detail.
Let me know what works well for you, what doesn’t, what would you like to see added, changed, or removed from the model?
If you enjoy the project and want to support future work, tips are always appreciated—they help me stay warm, and sane.
Cheers
Description
Even more training, and the last run for this dataset
FAQ
Comments (18)
which subs?
synthetic dataset = bad quality
can't use real people. i seen another lora that claims to use real photos from reddit. that isn't allowed. You can't use peoples likeness without permission
can't use real people. i seen another lora that claims to use real photos from reddit. that isn't allowed. You can't use peoples likeness without permission
@Santaonholidays Very reddit
@someone88 Two comments above with the exact same text… suspicious.
@Vyxen808 xD
@Vyxen808 never heard of copy and paste? clearly OP agreed. and it's common sense. if your in USA or especially state i am from. It's illegal to make any AI of any real people without permission and will result in jail time. It doesn't even need to be explicit. my state will put you in jail just for making an image in someone else likeness. There are some other countries who also will do the same like South Korea for example.
@zengrath Pretty sure korea would give you a d sentence lmao
@zengrath
Does the fact that some countries have such abhorrent laws make training on synthetic data any better? If you can't train on real data for fear of prosecution, let those whose government isn't quite fucked up enough do it.
@Unicom Him making LoRA on synthetic dataset doesn't prevent the other to do what they want so let him do what he wants and move on.
@juanitaz Certainly. Just don't present this as an advantage. This is a clear drawback that everyone who works with this understands, which is actually reflected in this thread.
@Unicom "Just don't present this as an advantage."
He didn't, people just explained why they would use a synthetic dataset.
You are the only one talking about "better" or "advantage".
Also you can make great realistic character LoRA based on synthetic dataset, anyone that did it seriously knows it so there is more subtleties than just "bad quality". Worse or not ideal is not necessarily equal to bad or unusable.
@juanitaz Synthetic data is always worse when it comes to realism. And this model is precisely about realism. I don't think there's anything else to discuss here.
So what's exactly the function of the Lora?,tried a simple naked image and even with the Lora the nipples were all deformed
If you are on Comfy, try using the Iterative Upscale nodes
@ja004311 Ok but how connect a the lora to upscaler?
Did you train this locally? If so, what trainer did you use and what are your hardware specs?