![Amateur Photography [Flux Dev] v2.0 Test](https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cdfe5617-e6dc-4ed5-be7f-b7015f86edc7/width=1320/03_Cosplayers_1.00_00001_ (2).jpeg)
Amateur Photography [Flux Dev] v2.0 is a significant improvement from v1.0
All images were generated with the following settings:
Flux_dev
t5xxl_fp8
Random Seed
896 x 1152
euler
20 steps
Prompt: Amateur photo of a Latino wearing a yellow t-shirt taking a mirror selfie in a dimly lit bathroom. There are some jpeg artifacts.
LoRA Weight: 0.00
LoRA Weight: 0.55
LoRA Weight: 0.75
LoRA Weight: 0.85
LoRA Weight: 1.00
For some reason the LoRA doesn't affect the particular prompt above to a large degree. I believe this is because Flux Dev already produces an amateur photo with this particular prompt. But regular Flux is unwieldy when it comes to getting the style you want -- so let's see what happens with a prompt that produces more professional looking photos (and more obviously AI)
Prompt: Amateur photo of two gorgeous blonde 20-year-old cosplayers posing for a photo at a Comic-Con convention. There is a crowd behind them.
LoRA Weight: 0.00
Pretty clearly AI and not amateur looking at all without the LoRA.
LoRA Weight: 0.55
LoRA Weight: 0.75
LoRA Weight: 0.85
LoRA Weight: 1.00
At all weights, the LoRA creates significantly more photo-realistic images than unaltered Flux Dev and it is because of the imperfections added to the image that makes things look real. I will test now with a more complicated prompt:
Prompt: Amateur photo from a 2007 blog of two hippies sitting on a bed. One of the hippies is male and the other is female. They are wearing colorful socks. There are psychedelic themed posters on the wall and the room is cluttered and there is smoke in the air. They look happy and tired. The room is dimly lit and there are some jpeg artifacts.
LoRA Weight: 0.00
LoRA Weight: 0.55
LoRA Weight: 0.75
LoRA Weight: 0.85
LoRA Weight: 1.00
The first generations without a LoRA don't look too bad until you see the generations with the LoRA. It's hard to believe it's AI generated until you see the anatomy issues.
Conclusion:
If you are trying for realism and regular flux isn't working, this LoRA will likely help with it. Different prompts get varying results. I recommend a weight around 0.6 - 0.8. The cons are that it might be harder to get subject prompt adherence and anatomy. There may be some workarounds that I haven't tested such as:
Generating with raw Flux and then using the LoRA with an img2img workflow
Generating the first 5-10 steps with raw Flux and then adding the LoRA for the remaining steps
ControlNet or IP adapter when and if a good workflow is available for Flux