Loading
Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
GENOVA APEX is an experimental model.
Model source
Source description
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
Sources
1 sourceModel artifacts
1 artifactgenovaAPEX_s.safetensors
safetensors · FP8 · FULL · 11.1 GB · SHA-256 c46b0346edc1…1f46 · Civitai
DownloadSource excerpts
1 excerptSource context: 2826 downloads · Type Checkpoint · Base model Flux.1 D
Sampling Steps : 8 steps
I use Hires. Fix because it provides significantly better results: 8 steps, upscale 1.2.
My configuration: RAM 64Gb, 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-gen P.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...
