New Changes in v2:

1. Adjusted the dataset (note that you may still see some bias towards white people but i suggest to prompt what you want and not say "woman" or "man")

2. Tagged the race, ethnicity and also physical attributes of the subjects so it should control the biasing towards plus-size people

3. Training dataset captions are now ~200 words per image (instead of 45-70 in v1). T5XXL is no joke lol. That means it can generate complex scenes, you can also position people and objects where you want (Base model can already do this, this lora just adds the realism and clutter to it). It may or may not work, so you can do some experimentation

4. It can also generate some high quality background blur pictures if you are into it. Prompt it using "cinematic feel" at weight 0.5 or 0.6 or other words that work for you

5. I may have messed up fingers in v1. I think v2 corrects this (if the image of base model have bad fingers, this lora tends to follow it)

6. Realism kicks in at weight's between "0.5-0.6". If you want to stay close to the output the base model generates without this lora, i suggest to stay between "0.5-0.6". Maximum realism is between 0.8 and 1.0. But be prepared to see some horrors lol (you can experiment yourself. These are just my observations based on my limited testing)

I would like to thank all the people who donated buzz for v1 version. I was skeptical at first to buy 5$ buzz and now I was able to train 6 Flux lora's in civitai. BTW 400 epoch v3 is cooking. Lets see how it goes haha