<h3 id="features"><span style="color:rgb(250, 176, 5)">Features</span></h3><ul><li><p>Cinematography Style and Theme from BladeRunner 2042</p></li><li><p>Yellow, Dark, and Orange colored themes</p></li><li><p>Compatible with both T2V &amp; I2V Wan2.2 Models</p></li><li><p>This LoRA was designed to provide a subtle influence on background, clothing, and scenes</p></li></ul><h3 id="tips"><span style="color:rgb(230, 73, 128)">Tips</span></h3><ul><li><p>Play around with LoRA strength below 1.0 if the LoRA is not 'working'.</p></li><li><p>SPEED LoRAs destroy dark scenes. I was only able to get really dark scenes by not using the popular low step speed LoRAs.</p></li><li><p>HIGHLY recommend reviewing and understanding how to correctly split steps for Wan2.2 models: <a target="_blank" rel="ugc" href="https://www.reddit.com/r/StableDiffusion/comments/1mkv9c6/wan22_schedulers_steps_shift_and_noise/?tl=fr">https://www.reddit.com/r/StableDiffusion/comments/1mkv9c6/wan22_schedulers_steps_shift_and_noise/?tl=fr</a></p></li><li><p>Workflow that correctly splits High / Low noise steps: <a target="_blank" rel="ugc" href="https://github.com/stduhpf/ComfyUI-WanMoeKSampler">https://github.com/stduhpf/ComfyUI-WanMoeKSampler</a></p></li><li><p>Example prompts. These prompts were used during training for the image/video text captions '<span style="color:rgb(130, 201, 30)">The scene is shrouded in darkness, with only the occasional glow from distant lights piercing the fog</span>', '<span style="color:rgb(130, 201, 30)">a dimly lit room, surrounded by various objects such as a desk, a table, a chair, and a lamp</span>', '<span style="color:rgb(130, 201, 30)">a man in a long coat standing in a dusty, yellow-toned environment. He appears to be observing something off-camera</span>'</p></li><li><p>Use negative prompts to remove items that keep appearing in the scene. For example, 'tires' when trying to generate a futuristic car that keeps having tires on it! lol</p></li></ul><h3 id="my-wan2.2-cover-videos"><span style="color:rgb(76, 110, 245)">My Wan2.2 Cover Videos</span></h3><ul><li><p>These videos have the ComfyUI metadata saved in them.</p></li></ul><h3 id="training"><span style="color:rgb(64, 192, 87)">Training</span></h3><p>Dataset images and video contained scenes without people facing the camera because of Civitai's real person terms of service policy.</p><p>I'm testing a fork of musubi-tuner that allows me to run validation tests against a dataset that is not in the training set. This should prevent over-fitting during training. Why validation testing is important: <a target="_blank" rel="ugc" href="https://github.com/spacepxl/demystifying-sd-finetuning">https://github.com/spacepxl/demystifying-sd-finetuning</a></p><h3></h3><p></p>

Files with SHA256: bb3d6af7d3a5364b793a12c80072b76ab79467b21044458733394c0a80d34c50