This workflow is an optimized iteration of the original WAN22 RAPID All-in-One 4-Steps NSFW with CineScale/High-Res model created by FaboroHacks [https://civarchive.com/models/1973876/rapid-wan22-all-in-one-mega-nsfwmotion-transfer].

The core addition lies in integrating new nodes for new SeedVR Upscaler and interpolation which is now really great for superior video upscaling and a Dual QWEN Analyst system powered by QWEN 3 VL (you'll need to download abliterated models if you wanna do gentelmany stuff), which analyzes both input videos and images to generate context-aware prompts. Also, I've integrated purging of VRAM and RAM everytime memory intensive operation is finished, because in original workflow I was getting stuck and OOM when trying to upscale, interpolate or do another run. Also, added sage attention optimization with Kijaji node.

This version is designed for users with mid-to-high-end hardware, such as NVIDIA RTX 3090 GPUs. 7-second input clip processes in approximately 20 minutes, depending on resolution and complexity. If VRAM issues arise on lower-end GPUs, reduce batch size or clip length. I already set up SeedVR to offload maximum amount of blocks, but it may be stuck when encoding and decoding VAE if you're low on VRAM. So make sure to adjust tile sizes as well.

A huge shoutout to the original creator FaboroHacks! Make sure to follow and support them!

About abliterated version of QWEN
You will download this (https://pastebin.com/DTssQDs4) name it as custom_models and save it as .json (custom_models.json)
You will put it in \custom_nodes\ComfyUI-QwenVL
Restart Comfy and QwenVL node will now have abliterated versions that will be downloaded automatically.