Credits: aramintastudio

Current SOTA model in my experiment:

Flux model: Flux1-A1

SDXL Base model: Gv4 is the most balanced model allowing both realistic and styled NSFW and SFW images. Better aesthetic than Fv5.

SDXL photorealist (SFW and NSFW) model: Fv5 is the way to go for hyper-realism including realistic NSFW images but it mostly lacks the styling capabilities of Gv2.

SDXL Illustration : Gv4 (SFW and NSFW). Cv6 is however still worth a try if you are not into NSFW images.

Flux A serie is my first Flux.1 model created by merging Flux-dev-fp8 with several Loras I have trained using my dataset. At this point it has to be considered as a WIP and it is not clear whether it will be possible to create a versatile base model using this approach. But Flux is obviously the SD3 we were all hoping for and its capabilities out of the box are quite amazing.

Image Generation Settings for SDXL models

DPM++ 2/3M SDE / Karras or Exponential are always a good bet with 25+ steps and CFG around 5-7. But DPM++ SDE / Karras with less steps (e.g. 12) and higher CFG (8-11) is worth a try.

The default CLIP Skip of 2 is also a good bet, but using 1 or 3-4 is also worth trying: 1 push more towards prompt adherence and 3-4 give sometimes a better result than the default focusing more on the "concepts".

Image Generation Settings for Flux models

My preferred settings are DPM++ 2M / beta or sgm_uniform or DDEIS / normal for the sampler / scheduler, beta giving a bolder stronger image. For a more subtle image, Euler / simple or beta seems a good bet.

CFG seems to have a huge impact on the final image and be very sensitive even to small variations.

For photos, CFG should remain low (1.5-2.5) to avoid plastic skin.

For fine art and illustration it is more complicated because it depends on the medium. For "rough" styles (painting, watercolours etc.), CFG should stay quite low in the 1.5-2.5 range but for anime or comic style, CFG needs often to be pushed further to achieve the desired style (3-6 or more).

If the image is messy/malformed or blurred, it is often because the CFG/steps are inappropriate for this image, but it is not always easy to know whether CFG/steps must be increased or decreased (at least to me 😊).

There is for sure a lot to learn concerning Flux behaviour which is quite different than SDXL and we will need to adapt.

Araminta Experiment - FV6