It's basically video to video. But instead of denoise we just skipped steps (so its refine).

Of course its using Lightx2v lora - Wan2.2-T2V-A14B-4steps-lora-250928 but any lightx2v lora will work good since the movement issue it not problem here.

You can use 1 or 2 steps, depending how much stuff you want to fix. Final step also affects changes - bigger step from 4, smaller changes (checkout the horses example).

- step 3,4: big changes
- step 5: some changes
- step 9,10 small changes
- step 10: very small changes
etc

Of course with 2 steps KSampler works 2x longer

• examples: upscaled on left, at top left there's info about steps used

• examples (2 videos at once) are downscaled because of ram problems with rendering both

• if there's some stuff normal T2V cant do (like a dick) add some T2V lora

• prompt is optional, but its nice to add very brief tags what's in the video

• at the final sharpening stage video is downscaled to make sure upscaler don't upscale blur

• mine 4090 24gb can upscale video up to 1megapixel 120 frames in ~300 sec

• workflow is using ⚡ Rife Tensorrt, super highly recommended, it's incredibly fast compared to regular (like 10x) frames interpolation, I'm attaching bat script I've used inside to run tensor file creation (for comfy portable). Place it in ComfyUI-Rife-Tensorrt folder (download rife49_ensemble_True_scale_1_sim.onnx to "ComfyUI-Rife-Tensorrt/models" folder first)

• monitor memory, if anything leaks into shared vram, close comfy, change resolution to smaller and try again

upscaler link - 2x-NomosUni_span_multijpg - OpenModelDB