Real HDR
This training used full raw images (RAW format) converted to 48bit png.
FLUX was then trained on the full range of 65k per channel color normalized to float.
Description
FAQ
Comments (11)
Not working, ; destroys face character
I have tested with character loras, in my testing the addition of the HDR lora did not cause character failure but rather the seed. In other words the character lora failed on that seed.
Although I am sure you could find a seed that the HDR caused a failure +/- I did not find this to be the case in general
"Sorry to hear you're having trouble. 'Destroying' a face is usually a sign of LoRA interference.
In my testing, I've found that because this is an HDR LoRA focused on high-detail texture, it can sometimes compete with Character LoRAs—especially highly realistic ones. Since both models are trying to influence the same 'noise patterns' (like skin texture and fine lines), the model can 'snap' to the wrong details, causing the character to shift.
A few tips to fix this:
Lower the weight: Try running this LoRA at 0.3 – 0.6 when using it with a character.
Prompting: Use more specific facial prompts to 'anchor' the character.
Order: Try swapping the order of the LoRAs in your prompt.
AI generation is rarely 'plug and play' with multiple LoRAs, but with a little weight adjustment, they should work together!"
PNG can only be 16 bit, not 48 bit?
16bits per channel 16x3
Had some more time to test the lora today. It is gorgeous. Im always after some off-lable use and here is where the magic happens. If you want some motion blur artistic light effects, this lora is the key to open that door (negative values used!). If you're into neon-noir-cyberpunk-style...I posted some pictures in the gallery. I also posted a video to show how sharp quality videos you get out of the rendered high detailed pictures. It is a must have lora for all kind of usages. Thx for sharing.
Thanks - Glad you found use cases for both directions on the LORA