<h2 id="consider-following-my-work-on-patreon-where-the-free-tier-unlocks-everything.-models-workflows-research-process."><a target="_blank" rel="ugc" href="https://www.patreon.com/CalvinHerbst">Consider following my work on Patreon, where the free tier unlocks everything. Models, Workflows, Research, Process.</a></h2><a target="_blank" rel="ugc" href="https://www.patreon.com/posts/flux2-klein-lora-149190475?utm_medium=clipboard_copy&amp;utm_source=copyLink&amp;utm_campaign=postshare_creator&amp;utm_content=join_link">Watch a deep dive of the training process here</a>A year ago, I released version 3, and was surprised to see the volume of both support and criticism. I still stand by my belief that we can not only take control of the technology's potential by training on our own imagery, but also that we can bring an empowering version of the post-AI-visual-sphere into realization through publishing tools made by individuals, not just corporations, and that are accessible to anyone with a laptop, not just those with industy credentials or formal education. The democratic nature of free, open-source tools will inherently create a lot of slop, but I feel expanding the reach of any medium is a net positive.You can support me on Patreon and get everything for free :)<a target="_blank" rel="ugc" href="https://www.patreon.com/c/CalvinHerbst">https://www.patreon.com/CalvinHerbst</a>Aesthetic Properties of the model: HerbstPhoto_v4_Flux2 produces intensely imperfect images that feel candid and alive. The model creates analog degradation micro-textures that break past the plastic look by introducing filmic softness, emulsion bloom &amp; hailation, optical artifacts - such as lens flares, light leaks, chromatic aberration, barrel distortion - and grain that behaves naturally across exposure levels. Compositions are moody and take form in chiaroscuro light, with dark regions that blanket the frame to create asymmetry and bright slivers that form hotspots to maintain balance. The contrast curve is aggressively low latitude, embracing clipped highlights and crushed shadows, while preserving a high black point to feel true to the celluloid nature of the images the model was trained on. Version 4 is trained for Flux 2 Dev from @Black Forest Labs because I beleive it’s the best image diffusion model, however it’s a heavy and can take several minutes to generate a single high-res image, so I will also be releasing an updated version for Z-image, Flux 1 Dev, and SDXL in the coming weeks for those who are looking to use less compute or create faster. Best Practices using the model:<ul><li>Prompts: Include “HerbstPhoto” in the prompt. Though the Flux 2 Model can handle prompts that are long and complex thanks to its incorporation of the minstral_3_small_fp8 text encoder from @Minstral AI I tuned this LoRA to produce dramatic effects even with simple language writing that does not include style, texture, and lighting tokens. </li><li>LoRA strength: 0.4 - 0.75. (0.73 sweet spot) 0.8-1.0 for less prompt adherence and max image texture/degradation. </li><li>Resolution: 2048x1152, though the model also produces good results across aspect ratios and sizes up to 2k. </li><li>Schedulers and Samplers: I tested every combination of Schedulers and Samplers for Flux 2 (378 total) and can recommend a handful of combinations that I tested on a Pro 6000 WK GPU @ 1024x1024 @ 20 steps that each have different aesthetics and render speeds. </li></ul><ol><li>dpmpp_2s_ancestarl + sgm_uniform: Best balance of texture &amp; fidelity. 160 sec. Render </li><li>er_sde + ddim_uniform: Good balance of texture &amp; fidelity. 60 sec. render</li><li>dpmpp_sde + simple: Softer focus, lower contrast, less artifacts, brighter. 130 sec. Render</li><li>dpmpp_3m_sde_gpu + simple: higher contrast, brighter, more chromatic aberrations. 60 sec. render</li><li>Ipndm + simple: Higher clarity, less softness, fewer artifacts, cooler. 60 sec. render</li><li>dpmpp_sde + ddim_uni: higher saturation, color shifting. 130 sec. Render  </li></ol>Training Process Overview:I used AI Toolkit from Ostris on an H200 GPU cluster from Runpod to train over 100 versions of the model, all using the same dataset + simple captions. For each run, I changed one parameter to get a clean A/B tests and figure out what actually moves the needle. I’ll share the full research soon :) After lots of testing, I am happy to finally release HerbstPhoto_v4_Flux2.Coming soon:HerbstPhoto_v4.1_Flux1DevHerbstPhoto_v4.2_ZImageHerbstPhoto_v4.3_SDXLHerbstPhoto_v4.4_Flux2_DarkAbyssHerbstPhoto_v4.5_Flux2_FishEyeHerbstPhoto_v4.6_Qwen_ImageEnhancer

Herbst Photo - analog film

HerbstPhoto_v4_Flux2.safetensors

HerbsPhoto_v4_Flux2

HerbstPhoto_v4_9b_fid

Files with SHA256: b55b55c6bd5d982492e649bb019edc9497c6dbb6e73b8bd36c61273085073878