Downloads available at: https://civitai.com/user/diffusional
V6: Improved dataset, diffusers training back end used via SimpleTuner, weight tuned with custom script.
Use simple prompts - "An old man in a wooden fishing boat",
Use Lora strength 1.0, but experiment with 0.2 - 2.0 as needed. Has a wide weight range
dies sampler+simple scheduler or Euler sampler+beta/simple scheduler (others may also work well)
Guidance 2 - 5 (usually 3.5), Steps 20 a 50 (usually 25 or 35)
Only add 'realistic' or 'fine hairs, visible pores, etc.' if you're not getting results
V5:
Improved dataset, trained for flux-dev on the flux-dev2pro fine-tune designed to improve training by pushing the model to work with smaller steps.
***Note: In my demo photos I had the file named 'flux-photo_v6.0_turbo' but it's the same file, there is no v6_turbo. I went through several versions and canned all but this one that stood out. Turbo? You might get away with a few less steps, but I didn't want to put turbo in the name.
V4-compatible/standard/extreme:
standard is usually the best, extreme with weight scaled down can be nice for exaggerated or forced photo realism, but can eat styles and turn things into normal people. compatible is pretty nuanced, could be nice with a higher weight.
V3-high-strength:
excellent on it's own with no prompt input about realism
V3-medium-strength:
excellent added to prompts that already have other loras or language about photorealism so it doesn't overbake the generation
flux lora trained on high quality photography images to produce realistic high quality photos without needing to add any tags or keywords (i.e. no 'a photo of' or 'realistic' etc.)
