<h1 id="goonwan-2.2-5b-is-a-nsfw-furry-finetune-of-wan-2.2-5b."><span style="color:rgb(50, 217, 86)"><strong>GoonWan 2.2 5B is a NSFW Furry finetune of Wan 2.2 5B.</strong></span></h1><p></p><h2 id="the-model-was-created-by-training-on-a-large-corpus-of-furry-and-non-furry-nsfw-videos-then-performing-svd-based-distillation-to-distill-the-wan-2.2-14b-high-and-low-models-into-the-5b."><span style="color:rgb(17, 127, 212)">The model was created by training on a large corpus of Furry and non-Furry NSFW videos then performing SVD based distillation to distill the Wan 2.2 14B high and low models into the 5B.</span></h2><p></p><h2 id="the-model-was-then-put-through-a-custom-svd-based-rl-training-loop-training-on-11k-videos-to-enhance-the-motion-and-temporal-consistency."><span style="color:rgb(255, 0, 247)">The model was then put through a custom SVD based RL training loop training on 11k videos to enhance the motion and temporal consistency.</span></h2><p></p><h2 id="the-wan-2.2-14b-mps-and-hps-reward-loras-were-then-converted-to-work-with-the-5b-using-svd-and-orthogonal-procrustes-alignment-and-then-merged-with-the-model."><span style="color:rgb(0, 255, 21)">The Wan 2.2 14B MPS and HPS reward Loras were then converted to work with the 5B using SVD and Orthogonal Procrustes alignment and then merged with the model.</span></h2><h2 id="edit:-the-wan-5b-turbo-lora-was-also-merged-with-the-model.-forgot-to-add-this-when-the-model-was-posted.-due-to-this-sometimes-you-can-get-slow-motion-outputs.-using-the-wannag-node-helps-alot-with-prompt-following."><span style="color:rgb(250, 176, 5)">EDIT: The Wan 5b turbo lora was also merged with the model. Forgot to add this when the model was posted. Due to this sometimes you can get slow motion outputs. Using the WanNAG node helps alot with prompt following.</span></h2><h2 id="the-model-performs-best-with-10-sampling-steps-euler-sampler-beta-scheduler-and-a-cfg-of-1-or-10-sampling-steps-and-a-cfg-of-2.0-2.5-going-above-a-2.5-cfg-can-cause-artifacts.-due-to-the-model-functioning-optimally-at-a-lower-cfg-and-lower-steps-it-generates-much-much-quicker-than-the-base-model."><span style="color:rgb(0, 255, 251)">The model performs best with 10 sampling steps, Euler sampler, Beta scheduler and a CFG of 1 or 10 sampling steps and a CFG of 2.0-2.5 going above a 2.5 CFG can cause artifacts. Due to the model functioning optimally at a lower CFG and lower steps it generates much, much quicker than the base model.</span></h2><p></p><h1 id="all-characters-are-18+.-this-is-intended-for-mature-audiences-only-you-and-you-alone-are-responsible-for-what-you-generate.-this-model-is-not-to-be-used-on-any-real-person!-18+-only!"><span style="color:rgb(250, 82, 82)"><strong>All characters are 18+. This is intended for mature audiences only you and you alone are responsible for what you generate. THIS MODEL IS NOT TO BE USED ON ANY REAL PERSON! 18+ ONLY!</strong></span></h1><p></p>

Files with SHA256: d1e2c5ff4f6265b782fb8187f749c4b53192e92ec2741d0876265438137a59cc