This finetune aims to improve reliability in realistic/photo-based styles while preserving <a target="_blank" rel="ugc" href="https://civarchive.com/models/1330309/chroma">Chroma</a>’s broad concept knowledge. The v1.3 flash version has the rank-256 lora (<a target="_blank" rel="ugc" href="https://civarchive.com/models/2032955?modelVersionId=2300817">from here</a>) baked in. bf16 and fp8 available ---&gt;. <a target="_blank" rel="ugc" href="https://huggingface.co/dawncreates/UnCanny-Photorealism-Chroma-GGUF">GGUFs on HuggingFace</a>.Prompting: Simple prompts describing what you want to see in natural language works well (<a target="_blank" rel="ugc" href="https://civarchive.com/images/114425805">example1</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/images/114426208">example2</a>). Chroma prompts work well. Examples of captioning style used in training: <a target="_blank" rel="ugc" href="https://civarchive.com/images/114430993">woman sitting</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/images/114425797">waterfall</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/images/114426216">wolf</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/images/114426205">woman in dessert</a>. Negative prompts don't work at CFG one, but above one negative prompts can be important.Example settings (not necessarily optimal) using <a target="_blank" rel="ugc" href="https://comfyanonymous.github.io/ComfyUI_examples/chroma/">Comfy's default Chroma workflow</a>: Base: Steps:: ~30-40. CFG: ~3.5 (best settings depend on steps/CFG/sampler, etc.). Flash (lora rank-128 or 256): Steps: 15-17. CFG 1. (Depends on lora rank, sampler, etc.) Sampler: Examples use <code>res_2m</code>, <code>dpmpp_sde</code>, or <code>exp_heun_2_x0_sde</code>. Others are also good. Scheduler: I like <code>bong_tangent</code> | <code>beta</code> &amp; <code>beta57</code> and others are also goodSupport: Have too much money? Want to support further training? <a target="_blank" rel="ugc" href="https://ko-fi.com/dawncreates">https://ko-fi.com/dawncreates</a>Training Details The model was trained locally, using Chroma-HD as the base. Each epoch included images at 3–5 different resolutions, though only a subset of the dataset was used per epoch. Except for the extra resolutions, OneTrainer's default config for 24gb Chroma finetuning was used. The dataset consists almost exclusively of SFW-images of people and landscapes, so to retain Chroma-HD's original conceptual understanding, several layers were merged back at various ratios. All the juice, compositions, subjects, and concepts come from Chroma itself, my model just nudges it towards realism. Honestly, this version is a showcase of how good Chroma is. So get to work on Chroma finetuners - it has so much potential!All images were captioned using JoyCaption: <a target="_blank" rel="ugc" href="https://github.com/fpgaminer/joycaption">https://github.com/fpgaminer/joycaption</a>The model was trained using OneTrainer: <a target="_blank" rel="ugc" href="https://github.com/Nerogar/OneTrainer">https://github.com/Nerogar/OneTrainer</a>

UnCanny (Photorealism Chroma)

Flash version of v1.2. Rank-128 of flash-heun baked in.

uncannyPhotorealism_v12Flash.safetensors

v1.2-flash

v1.3 - Flash

Files with SHA256: 35cc04df2bc881c40aae1efac99bd3fede14f324f42f785989f334e11f3f8639