CalvinHerbst/HerbstPhoto_v4_Flux.2-Klein-9b

HerbstPhoto_v4_Flux2_Klein_9B

Base Architecture: Flux2-Klein-9B (Black Forest Labs)

Consider following my work on Patreon (free tier unlocks everything): https://patreon.com/CalvinHerbst

About

HerbstPhoto_v4 is a LoRA fine-tuned on a body of personal analog photography to produce intensely filmic images that break past the synthetic, plastic look common in AI generated imagery. The model introduces analog micro-textures, emulsion bloom, halation, optical artifacts such as lens flares, light leaks, chromatic aberration, and barrel distortion, along with grain that behaves naturally across exposure levels. The contrast curve is aggressively low latitude, embracing clipped highlights and crushed shadows.

This release targets Flux2 Klein 9B specifically because it runs on consumer hardware. Accessibility has been a priority: these tools should be available to anyone with a laptop.

All training data is original analog photography owned by the creator.

Two Versions

v4_Texture Heavier grain, higher contrast, more highlight bloom, increased soft focus, under-exposed images, and more frequent optical artifacts like lens flares and light leaks. The destructive end of the aesthetic.

v4_Fidelity Retains more subject visibility with milder changes in grain, contrast, bloom, focus, exposure, and artifacts. A controlled burn that preserves detail while still eroding the clinical digital look.

Recommended Settings

Base Model: flux2-klein-base-9b-fp8 (The base-fp8 version has better textures than the standard klein-9b, though the non-fp8 version has better fidelity across seeds)

Prompt: Include the tag "HerbstPhoto"

Image Resolution: 1344x768 (Works well from 1024x576 up to 2304x1156)

LoRA Strength: 0.73 is the sweet spot (Range: 0.4 to 1.0. Use 0.8 to 1.0 for less prompt adherence and maximum image texture/degradation)

Flux Guidance: 2.5 (Range: 2.1 to 2.9)

Scheduler and Sampler: dpmpp_2s_a + sgm_uni

Denoise: 1.00 (0.6 to 0.9 when using the LoRA for Image to Image)

Seed Note: The fp8 versions have higher seed dependency than the non-fp8 and dev models. It usually takes 5 to 10 runs to get a good seed with fp8.

Links

PATREON: https://patreon.com/CalvinHerbst WEBSITE: https://calvinherbst.com LINKEDIN: https://linkedin.com/in/calvinherbst TWITTER: https://x.com/calvinherbst1 INSTAGRAM: https://instagram.com/calvin_herbst