I basicaly used <a target="_blank" rel="ugc" href="https://civarchive.com/models/1390683/coyottes-moviemaker-wan-video-audio-generation">COYOTTE's MOVIEMAKER v3</a>, works nearly perfectly, but also I liked the idea of a second model pass from workflow <a target="_blank" rel="ugc" href="https://civarchive.com/models/1474890">Wan i2v with 720p smoothing v1.1</a>, because after the first pass there are artifacts on faces, etc, second pass helps.But my rig (4070 ti super 16 GB VRAM, 32 GB RAM) can't handle more then one generation (14B GGUF Q4_K_M + 1.3B) without full restarting ComfyUI. I had a VRAM overflow problem on the first ksampler, when i tried to generate second time.After many experiments i merged two workflows and simplify them as posible, which works nearly perfectly for me. With this solution i don't need even unload models or cache in the end.It's just experiment, might be useful to someone with similar problems.For Comfy I use "--lowvram" parameter and sage attention, Also try to add "disable_smart_memory = true" inside ..\ComfyUI\user\default\ComfyUI-Manager\config.iniif everything alright for me first generation around 700s, next around 500s.For VRAM economy could try to disable hardware acceleration in browser for generation process.Thanks to <a target="_blank" rel="ugc" href="https://civarchive.com/user/coyotte">Coyotte </a>and <a target="_blank" rel="ugc" href="https://civarchive.com/user/HazardAI">HazardAI.</a>

Files with SHA256: 059ca6cea5710cca04e801a6b5ddad7a827f903775cd31a51628042262609135