HerbstPhoto_v4_Flux2

A LoRA model for Flux 2 Dev trained exclusively on analog photography I own the rights to. Produces intensely imperfect images that feel candid and alive.

If you are interersted in following my releases, consider subscribing to my patreon, where the free tier unlocks everything: Patreon

Model Description

HerbstPhoto_v4_Flux2 breaks past the plastic look of AI-generated images by introducing authentic analog characteristics: filmic softness, emulsion bloom & halation, optical artifacts (lens flares, light leaks, chromatic aberration, barrel distortion), and grain that behaves naturally across exposure levels. The contrast curve is aggressively low latitude, embracing clipped highlights and crushed shadows.

This model represents my belief that we can take control of AI's potential by training on our own material—and that we can bring an empowering version of image generation into reality through tools made by individuals, accessible to anyone with a laptop.

Intended Use

Creative image generation with an authentic analog film aesthetic. Ideal for photographers, filmmakers, and artists looking to move beyond the sterile perfection of typical AI outputs.

Usage

Trigger Word

Include HerbstPhoto in your prompt.

LoRA Strength

Strength	Effect
0.4 - 0.75	Balanced analog aesthetic with good prompt adherence
0.73	Sweet spot
0.8 - 1.0	Maximum texture/degradation, reduced prompt adherence

Resolution

• Recommended: 2048x1152 (16:9) or 2488x2048
• Produces good results across aspect ratios and sizes up to 2K

Schedulers & Samplers

Tested every combination. These work best:

1. dpmpp_2s_a + sgm_uniform
2. er_sde + ddim_uniform
3. dpmpp_sde + simple
4. dpmpp_3m_sde_gpu + simple
5. ipndm + simple
6. dpmpp_sde + ddim_uni

Prompting

Flux 2's incorporation of the mistral_3_small_fp8 text encoder handles long, complex prompts well—but I tuned this LoRA to produce dramatic effects even with simple language. You don't need style, texture, or lighting tokens.

Training

• Base Model: Flux 2 Dev (Black Forest Labs)
• Training Framework: AI Toolkit (Ostris, LLC)
• Hardware: H200 GPU cluster via Runpod
• Methodology: 100+ training runs, changing one parameter per run for clean A/B testing
• Dataset: My own analog photography (full rights ownership)

Ethical Training

This model is trained exclusively on photographs I created and own. No scraped data, no unlicensed material.

Resources

• Model Download: Patreon (Free)
• ComfyUI Template & Tutorial: Link

Upcoming Releases

Versions for Flux 1 Dev, Z-image, and SDXL coming soon for faster generation and lower compute requirements.

Author

Calvin — Filmmaker and creative technologist Patreon

License

Please check the Patreon page for license terms.