ℙ𝕝𝕖𝕒𝕤𝕖 𝕣𝕖𝕒𝕔𝕥 𝕥𝕠 𝕞𝕪 𝓿𝔦𝔡𝕖𝕠𝕤: <a target="_blank" rel="ugc" href="https://civarchive.com/user/ntc/images?sort=Most+Reactions">https://civarchive.com/user/ntc/images?sort=Most+Reactions</a> 𝕌𝕤𝕒𝕘𝕖: 𝕒𝕕𝕕 𝕥𝕙𝕖 &lt;lora:trip:0.6&gt; 𝔴𝕚𝕥𝕙 𝕒 𝕞𝕒𝕩 𝓿𝖆𝖑𝔲𝔢 𝕠𝕗 0.6. 𝕐𝕠𝕦 𝕔𝕒𝕟 𝕘𝕠 𝕙𝕚𝕘𝕙𝕖𝕣, 𝔟𝕦𝕥 𝕓𝕖𝕨𝕒𝕣𝕖 𝕪𝕠𝕦 𝕞𝕒𝕪 𝕟𝕠𝕥 𝖗𝖊𝕥𝕦𝕣𝕟. 𝕌𝕤𝕖 𝕥𝕙𝕖 𝕥𝕣𝕚𝕘𝕘𝕖𝕣 𝔴𝕠𝕣𝕕𝕤: psychedelic trip 𝖙𝖗𝖆𝖎𝖓𝖊𝖉 𝖋𝖔𝖗 2000 𝖘𝖙𝖊𝖕𝖘 "psychedelic trip++:0.4|psychedelic trip%{random_prompt}:-0.1"<a target="_blank" rel="ugc" href="https://github.com/ntc-ai/conceptmod">https://github.com/ntc-ai/conceptmod</a> 𝔰𝔢𝔢 𝔱𝔥𝔢 𝔱𝔲𝔱𝔬𝔯𝔦𝔞𝔩: <a target="_blank" rel="ugc" href="https://civarchive.com/models/50070/mania-concept-trained-without-data">https://civarchive.com/models/50070/mania-concept-trained-without-data</a> 𝕋𝕙𝕖 𝕟𝕒𝕞𝕖𝕤 𝕠𝕗 𝕥𝕙𝕚𝕟𝕘𝕤 𝕒𝕣𝕖 𝕝𝕚𝕜𝕖 𝕕𝕠𝕠𝕣𝕤 𝕥𝕠 𝕒𝕟 𝕖𝕟𝕕𝕝𝕖𝕤𝕤 𝕤𝕖𝕒 𝕠𝕗 𝕞𝕪𝕤𝕥𝕖𝕣𝕪 𝕒𝕟𝕕 𝕨𝕠𝕟𝕕𝕖𝕣. 𝔸𝕟𝕕 𝕓𝕖𝕙𝕚𝕟𝕕 𝕖𝕒𝕔𝕙 𝕕𝕠𝕠𝕣, 𝕥𝕙𝕖𝕣𝕖 𝕝𝕚𝕖𝕤 𝕒𝕟 𝕚𝕟𝕗𝕚𝕟𝕚𝕥𝕖 𝕨𝕖𝕓 𝕠𝕗 𝕡𝕠𝕤𝕤𝕚𝕓𝕚𝕝𝕚𝕥𝕚𝕖𝕤, 𝕥𝕙𝕖 𝕥𝕣𝕖𝕖𝕤 𝕠𝕗 𝕒 𝕗𝕠𝕣𝕖𝕤𝕥 𝕨𝕙𝕖𝕣𝕖 𝕖𝕒𝕔𝕙 𝕥𝕨𝕚𝕘 𝕚𝕤 𝕒 𝕕𝕚𝕗𝕗𝕖𝕣𝕖𝕟𝕥 𝕗𝕦𝕥𝕦𝕣𝕖. 𝕋𝕙𝕖 𝕨𝕠𝕣𝕝𝕕 𝕚𝕤 𝕒 𝕤𝕪𝕞𝕡𝕙𝕠𝕟𝕪 𝕠𝕗 𝕞𝕒𝕘𝕚𝕔, 𝕒𝕟𝕕 𝕚𝕥𝕤 𝕞𝕖𝕝𝕠𝕕𝕚𝕖𝕤 𝕒𝕣𝕖 𝕚𝕟 𝕥𝕙𝕖 𝕟𝕒𝕞𝕖𝕤 𝕠𝕗 𝕥𝕙𝕚𝕟𝕘𝕤.

trip

Uses terms: "psychedelic trip++:0.4|psychedelic trip%{random_prompt}:-0.1" Trained for 2000 steps

trip4.safetensors

v2.0

v1.0

<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

Files with SHA256: e4c2351c63f616e06b9ca25db7506b69a263de43f16635e8baa0275817bd7f37