GENOVA APEX is an experimental model.

I created by merging my previous Gene Quantum model https://civitai.com/models/885573/flux-gene-beauty with DEDISTILLED models and my own trained LoRAs. The process involved multiple stages and is purely an experiment to achieve high detail and realism similar to DEDISTILLED in just 8 generation steps(+ 8 generation steps Hires. Fix) .

Version (UNet) – This version requires Flux VAE, Clip-L, and T5XXL to work effectively with the Flux development model.

Recommended Settings for Optimal Generation

• Sampling Method and Schedule Type: Euler Beta

• Sampling Steps: 8 steps

• I use Hires. Fix because it provides significantly better results: 8 steps, upscale 1.2.

• My configuration: RAM ??Gb, VRAM 16Gb

• Workflow for ComfyUi:
  I use a straightforward workflow for image generation, available here: Flux Photos DEV Workflow

  • HyperDetailer workflow FLUX PHOTOS DEV (HyperDetailer update) https://civitai.com/models/670083

  However, it requires a lot of memory because the process includes nodes for detail enhancement and upscaling, which also deliver the best results.

• Please note that the quality on Forge may be lower than on ComfyUI. Here are the parameters I used: last update Forge, steps 8, euler beta, Distilled CFG Scale 3.5, CFG Scale 1 and I recommend use Hires.Fix Upscaler Latent or other be better, 8 steps, Upscale by 1,2 (higher upscale factors give better results but are slower and require more VRAM/RAM), Denoising strength 0,7, Distilled CFG Scale 3.5, CFG Scale 1 .

• My Prompt Writing Assistant: https://civitai.com/articles/7868/gpts-flux-photo-prompt-genP.S. The results I observe during the modeling process are even better than those I obtain from the final output of this process. Therefore, I continue experimenting, as the reason for this discrepancy has not yet been found...