F23photo is a LoRA trained on a curated dataset of my own photographs taken with a Nikon D7500. Unlike a raw camera reproduction, this model is based on heavily processed, cinematic imagery. It focuses on a specific visual language: deep atmospheric tones, polished cinematic lighting, and the high-end organic texture of a well-tuned DSLR pipeline.
This LoRA is the result of an obsessive polishing process involving over 115 hours of iterative retraining and fine-tuning to reach this level of aesthetic consistency.
Flux.2dev
Currently the only base model.
v1.0 — 12 images. A light touch, the most stable version. Works best with 2mp, guidance 4-8.
v2.0 — 52 images. Lost.
v2.1 — 52 images. Relatively stable, though inconsistent. Works best with 640-1440p, guidance 2.5-4.
v2.2 (EXP) — 52 images, retrained in q6k with my fork. 640-1440p, guidance 2.5-8.
v2.3 (EXP) — 52 images, retrained in q8 with my fork. 640-1440p, guidance 2.5-8.
v2.5 — 57 images, retrained in q8 with my fork. The most balanced and photorealistic version to date. Note that this is still an intermediate release; it can be unstable, but when it hits, the result is exceptionally rewarding. Works best with 640-1440p, guidance 10.
This fork optimizes certain Flux.2dev training processes. Currently, only q4 TE override is supported.
Workflow
To get the most out of this LoRA, I recommend using my optimized workflow with a native upscaler specifically tuned for this model: here
I will be very happy to see your ideas that came to mind when exploring the possibilities of this LoRa, so feel free to share your generations! :>





