from the <a rel="ugc" href="https://civarchive.com/user/mdmachine">creator himself</a>:: super awesome : ) from chat below When I run this usually I use a chained (3) samplers in comfyui.This would use teacache set around .25. The CFG at 1.5 is most important on the 1st sampler as it helps define the initial noise a little more cohesively at the expense of some speed.for example @ 45 Steps:KSampler Adv #1: Add noise enabled, steps 45, cfg 1.5, sampler ipndm, scheduler kl_optimal, start at step 0, end step 15, return with leftover noise trueKsampler Adv #2: Add noise disabled, steps 45, cfg 1.5, sampler gradient_estimation, scheduler kl_optimal, ststart step 15, end step 30, return with noise trueKSampler Adv #3: Add noise disabled, steps 45, cfg 1.0, sampler_name res_multistep, scheduler kl_optimal, start step 30, end step 45, return with noise falseWorks well @ 24 steps using:ipndm_v &gt; euler &gt; dpmpp_2musing sgm_uniform for the scheduler.<hr />setting things up and uploading---tested and it is fast : )-downside-- possible the baked in detailers overriding added loraignore my prompt in the examples the rest of the data is fineupside is same steps with this look more polished then others with same stepfirst up is the Detail Plus! 8 step which i use at 12second 16 step which i use at 20 stepMerge-Detail.enhancement.and.Accelerationmerged models, built upon the base models:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/Freepik/flux.1-lite-8B-alpha">Freepik's Flux.1-Lite-8B-alpha</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/nyanko7/flux-dev-de-distill">Flux-dev-de-distill</a></li></ul>Detail enhancement and acceleration techniques have been applied, particularly optimized for NVIDIA 4XXX cards (maybe 3XXX too). The goal is to have high efficiency accelerated models with lower overhead.The (Detail Plus! De-Re-Distilled | Detail Plus! Distilled) models can be used with CFG remaining at 1-ish. Also the baked-in accelerators mostly work as intended. These are products of workflow optimizations, as such they can change/appear/disappear at any point and time.Detail Plus! - Built upon the base model <a target="_blank" rel="ugc" href="https://huggingface.co/Freepik/flux.1-lite-8B-alpha">Freepik's Flux.1-Lite-8B-alpha</a>:Detail Enhancement Used:<ul><li>Style LORA - Extreme Detailer for FLUX.1-dev (Weight: 0.5) (<a target="_blank" rel="ugc" href="https://civarchive.com/models/832683">Model Link</a>)</li><li>Best of Flux: Style Enhancing LoRA (Weight: 0.25) (<a target="_blank" rel="ugc" href="https://civarchive.com/models/821668">Model Link</a>)</li></ul>Acceleration Credits:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/ByteDance">ByteDance</a>, a leading AI technology company, optimized the following models:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/ByteDance/Hyper-SD/blob/main/Hyper-FLUX.1-dev-16steps-lora.safetensors">Hyper-FLUX.1-dev-16steps</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/ByteDance/Hyper-SD/blob/main/Hyper-FLUX.1-dev-8steps-lora.safetensors">Hyper-FLUX.1-dev-8steps</a></li></ul></li></ul>credit and info from <a target="_blank" rel="ugc" href="https://huggingface.co/mdmachine/FLUX.Model.Merge-Detail.Enhancement.and.Acceleration">mdmachine</a>

Files with SHA256: fee4b359c7c4f42dc328b4792585fcc93c672b244d1b5c06d8ad78c014919083