<h2 id="v2-update:-used-a-higher-rank-and-a-better-distillation-method-performance-should-be-much-better."><span style="color:rgb(190, 75, 219)">V2 Update: Used a higher rank and a better distillation method performance should be much better.</span></h2><h1 id="this-is-a-nsfw-lora-that-also-works-for-sfw-made-by-distilling-the-wan-14b-model-into-the-1.3b-model-and-extracting-it-as-a-lora-at-a-high-rank."><span style="color:rgb(190, 75, 219)">This is a NSFW lora that also works for SFW made by distilling the Wan 14b model into the 1.3b model and extracting it as a lora at a high rank.</span></h1><h3 id="you-will-see-&quot;lora-key-not-loaded&quot;-errors-in-the-console-this-is-normal-and-the-lora-will-work.-this-needs-to-be-used-with-other-nsfw-lorasnsfw-wan-models."><span style="color:rgb(64, 192, 87)">You will see "lora key not loaded" errors in the console; this is normal, and the lora will work. This needs to be used with other NSFW loras/NSFW Wan models.</span></h3><h2 id="this-works-best-applied-to-wan-fun-v1.1-inp-1.3b-as-its-a-much-better"><span style="color:rgb(64, 192, 87)">This works best applied to wan fun v1.1 inp 1.3b as its a much better</span></h2><h1 id="created-to-distill-wan-14b's-quality-and-motion-coherence-into-the-1.3b-for-quality-nsfw-generations.-you-should-notice-less-anatomical-glitching-increased-depth-and-realism-and-overall-motion-quality-and-stability.-make-sure-your-sampling-and-scheduler-settings-are-correct-or-else-you-will-never-get-a-good-wan-generation-due-to-insufficient-sampling.-on-a-rtx-3090-it-takes-58-seconds-to-make-a-512x512-33-frame-video-which-is-4x-faster-than-using-accvid-with-wan-14b.-you-should-notice-it-behaving-alot-like-the-14b-in-terms-of-quality.-genitals-will-prob-still-look-a-bit-off-since-wan-is-notorious-for-this."><span style="color:rgb(190, 75, 219)">Created to distill WAN 14B's quality and motion coherence into the 1.3b for quality NSFW generations. You should notice less anatomical glitching, increased depth and realism, and overall motion quality and stability. Make sure your sampling and scheduler settings are correct or else you will never get a good wan generation due to insufficient sampling. On a rtx 3090 it takes 58 seconds to make a 512x512 33 frame video which is 4x faster than using Accvid with wan 14b. You should notice it behaving alot like the 14b in terms of quality. Genitals will prob still look a bit off since Wan is notorious for this.</span></h1><h1 id="it-is-a-large-file-size-due-to-it-being-a-hybrid-rank-lora-this-is-required-to-capture-all-of-the-motion-coherence-and-quality-from-the-14b-model"><span style="color:rgb(64, 192, 87)">It is a large file size due to it being a hybrid rank lora, this is required to capture all of the motion coherence and quality from the 14b model</span></h1><h1 id="i-will-post-a-sfw-version-later-as-well"><span style="color:rgb(130, 201, 30)">I will post a SFW version later as well</span></h1><h1 id="the-large-file-size-does-not-increase-the-vram-requirements-to-use-it-since-it-is-a-lora."><span style="color:rgb(64, 192, 87)"><u>The large file size does not increase the VRAM requirements to use it since it is a lora.</u></span></h1><h3 id="this-lora-is-the-product-of-a-high-fidelity-distillation-from-a-wan-14b-teacher-model-created-using-a-normalization-svd-method-to-ensure-structural-purity.-it-employs-a-surgical-hybrid-rank-approach-where-capacity-is-allocated-based-on-layer-function:-a-full-1536-rank-is-applied-to-all-attention-(attn)-layers-for-maximum-coherence-and-a-384-rank-is-used-for-all-other-supporting-layers-for-stability."><span style="background-color:rgb(30, 30, 30);color:rgb(64, 192, 87);font-family:&quot;Google Sans Text&quot;, &quot;Helvetica Neue&quot;, sans-serif;font-size:14px"><strong>This LoRA is the product of a high-fidelity distillation from a WAN 14B teacher model, created using a Normalization-SVD method to ensure structural purity. It employs a surgical hybrid-rank approach where capacity is allocated based on layer function: a full 1536-rank is applied to all </strong></span><span style="color:rgb(64, 192, 87)"><strong>attention (</strong></span><span style="background-color:rgb(26, 28, 30);color:rgb(64, 192, 87);font-family:&quot;DM Mono&quot;, monospace;font-size:13px"><strong>attn</strong></span><span style="color:rgb(64, 192, 87)"><strong>) layers</strong></span><span style="background-color:rgb(30, 30, 30);color:rgb(64, 192, 87);font-family:&quot;Google Sans Text&quot;, &quot;Helvetica Neue&quot;, sans-serif;font-size:14px"><strong> for maximum coherence, and a 384-rank is used for all </strong></span><span style="color:rgb(64, 192, 87)"><strong>other supporting layers</strong></span><span style="background-color:rgb(30, 30, 30);color:rgb(64, 192, 87);font-family:&quot;Google Sans Text&quot;, &quot;Helvetica Neue&quot;, sans-serif;font-size:14px"><strong> for stability.</strong></span></h3><h1></h1><p></p>

Files with SHA256: e759424caf0ec622e9c98d4e28004c0d33075c0efeb2618ddbe627697e59d4ce