+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Check out my SDXL Models:
Stylized model: RAYBURN
Realistic model: RAYMNANTS
Painterly model: RAYCTIFIER
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

Introducing RAYFLUX PHOTOPLUS

PHOTOPLUS is born from my attempt at finetuning flux chins out of RAYFLUX. Street photography, Botanical garden, Zoo, and a fair share of synthetic data generated with my SDXL model, RAYMNANTS.

Hope you'll enjoy,

R.

HOW IS IT DIFFERENT?

PHOTOPLUS is a spinoff from RAYFLUX made to look more like actual photos. There are less Flux chins (not fully gone though), some improvement in anatomy (no major changes in NSFW), altho hands are better except when part of the subject explicitly, It's also pretty decent at lens flares, godrays, and kinds of bokeh.

In general, the pictures might look less color corrected, with more natural lighting. That also means that compared to RAYFLUXv1.0 it does lead to less cinematic lighting, less pizzaz if you want. Faces and skin textures are generally improved, as well as interiors or macro-style objects in general.

TL;DR: if you want images that look more like real photos PHOTOPLUS is your best choice, if you want more exotic, cinematic images, you can still use RAYFLUXv1.0.

EXAMPLES

1. Close up of the cataract blind eyes of a very old and wrinkly icelandic woman, 90 years old, white hairs, white eyebrows, marked skin, weathered rugged skin.
   
   

2. A vast expanse of grassland, two-dimensional, sunrise, intricate play of light and shadows, serene moments captured.
   

3. 1990s style analog flash photography.

   A young woman with short blonde hair and a futuristic bodysuit, holding a glowing energy weapon. She is in a sleek, high-tech spaceship cockpit, with holographic displays and stars visible through the large window. The style is clean sci-fi with a focus on advanced technology.

   It looks like it has been taken with a cheap disposable camera, with visible film grain and slight lens distortion.
   

SETTINGS

PHOTOPLUS is more fickle than RAYFLUXv1.0. I narrowed it down to 8 Samplers and 3 schedulers. Here's a few grids to illustrate those choices. I attached those grids to this post so you can zoom on them.
As you can see, some or sameish, some more different, some give you more noise but also more details, some are smoother. Depending on what you're trying to do, some might perform better than others, so don't hesitate to experiment a bit!

• Like RAYFLUX, keep both max shift and base shift at 1.0 and Flux conditioning stays at 3.5

• If you don't know which sampler/scheduler to use, just go for DPM2 Ancestral with Beta scheduler. As a reminder, RAYFLUXv1.0 used Heun/beta.

• For steps, Photoplus needs a bit more love. 14-20 steps will work but leave you with some dithering patterns in the noise, but nothing that can't be fixed with an upscaler. 24-30 steps seems ideal with most samplers.

• For the rest, keep it like the base RAYFLUX: load the model with the weight you want (I use fp8_e5m2, but e4m3fn or e4m3fn_fast work too) and my favorite encoders combo are the same: t5xxl_fp8_e4m3fn_scaled and ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.

I usually do a second pass with a SD upscale at 2 steps with denoise at 0.3. I switch the upscaling model depending on the style I'm looking for e.g. 1x_ITF_SkinDiffDetail_lite for grainy portraits (works great at x2 too!), 4xUltrasharp for generic illustration, 4x_Foolhardy_Remacri for softer, cleaner digital illustration style. The prompt you use on the SD upscale is very important with Flux in general, don't hesitate to abuse it to nudge the upscale in the direction you want (i.e. write grainy for more grain, soft for less grain, add details about your protagonist, etc.)

KNOWN ISSUES/QUIRKS

• Due to the way conditioning work, Flux's inherent strengths and weaknesses and my limited dataset (99% of my pictures are from Western Europe), PHOTOPLUS will underperform at ethnicity or mixing influences.
  
  For instance, prompting for a japanese woman will generally be okay, but blonde japanese woman will usually turn her into a caucasian woman. You will have the same issue with black skin, especially really dark skin tones. The same will also happen with old people when you start adding more details to the prompt.
  
  You can check this video from Matteo Spinelli that actually explains how this all happens (particularly the segment around 17.30).
  
  The fix will be usually crank up the weight (i.e. replace woman with black skin by woman with (black skin:2) for instance) or rephrase the prompt (i.e. use elderly instead of old, or use both)

  
  I apologize for this lack of support for varied ethnicity etc., I'm aware of the biases of my data and will work on improving this in the future.