<p>A streamlined image-to-video workflow built on Wan 2.2 14B MoE with dual high-noise/low-noise experts for crisp motion and stable detail from a single reference image. Positive/negative CLIP prompts condition motion planning in WanImageToVideo, then WanMoeKSampler blends HighNoise and LowNoise tracks for dynamic yet coherent denoising. The pipeline exports 16 FPS video natively, with optional RealESRGAN x2 upscaling and RIFE interpolation to 32 FPS for social-ready outputs.​</p><h2 id="key-features">Key features</h2><ul><li><p>Dual-expert sampling with HighNoise and LowNoise models via WanMoeKSampler for motion energy and detail stability.​</p></li><li><p>Automatic size passthrough from the reference image, avoiding aspect distortion and simplifying setup.​</p></li><li><p>Post stack: VAEDecode to frames, CreateVideo to MP4, optional RealESRGAN upscale, and RIFE x10 temporal interpolation.​</p></li></ul><h2 id="inputs">Inputs</h2><ul><li><p>Reference image loader, positive and negative CLIP prompts, optional audio passthrough in writer.​</p></li><li><p>Defaults: 960×960, length 65 frames, 16 FPS base.​</p></li></ul><h2 id="recommended-settings">Recommended settings</h2><ul><li><p>Sampler: UniPC + Beta scheduler, steps 4–5, denoise 1.0, boundary 0.5, sigma shift 5.​</p></li><li><p>LoRA: LightX2V HighNoise strength 5.6, LowNoise strength 2.0.​</p></li></ul><h2 id="where-each-model-goes">Where each model goes</h2><ul><li><p>Diffusion models (GGUF):</p><ul><li><p>Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf — loaded by LoaderGGUF and routed to the HighNoise branch. Place under models folder used by GGUF loader.​</p></li><li><p>Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf — loaded by LoaderGGUF and routed to the LowNoise branch. Same models folder as above.​</p></li></ul></li><li><p>VAE: pig_wan_vae_fp32-f16.gguf — loaded by VaeGGUF and connected to WanImageToVideo and VAEDecode. Place under models/vae.​</p></li><li><p>CLIP text encoder: cow-umt5xxl-q8_0.gguf — loaded by ClipLoaderGGUF with vision “wan” and projection “default”. Place under models/clip.​</p></li><li><p>LoRAs: wan_loras/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors — loaded by LoraLoaderModelOnly per branch, then into Power Lora Loader. Place under models/loras/wan_loras.​</p></li><li><p>Upscaler: RealESRGAN_x2plus.pth — via UpscaleModelLoader into ImageUpscaleWithModel. Place under models/upscale_models.​</p></li><li><p>Frame interpolation: rife49.pth — used by RIFE VFI node with factor 10 and scale 2. Place under models/rife.​</p></li></ul><h2 id="node-highlights">Node highlights</h2><ul><li><p>WanImageToVideo: prepares latents from image and prompts at defined width/height/length.​</p></li><li><p>WanMoeKSampler: blends high/low experts with CFG per-branch for quality control.​</p></li><li><p>CreateVideo: 16 FPS assembly; second writer used post-RIFE for 32 FPS export.​</p></li><li><p>Utility: multiple cleanGpuUsed and RAMCleanup nodes to stabilize long runs on limited VRAM.​</p></li></ul><p></p>

Files with SHA256: d3c1ecc5c31fbcd6738087ce4bffee581394299ee4ae155a4cc9e0939be93ce2