FAL Video2Video Request

This is the pack's heaviest video node, and it's built for a specific job: pose-driven motion transfer. Feed it a driving video (a source clip of someone moving) and a reference image (the subject you want performing that motion), and it calls FAL's fal-ai/wan-22-vace-fun-a14b/pose endpoint to generate a new video of your reference subject doing whatever the driving video does. VACE-style motion transfer is one of the more genuinely useful things the Wan ecosystem does well, and this node exposes essentially all of the endpoint's surface area rather than a trimmed-down subset.

The inputs that actually drive the result: video_name points at your driving video, already present in ComfyUI's input/ folder - not a URL. prompt describes the scene/subject; the node ships a long, opinionated default negative_prompt covering the usual video-generation failure modes (extra fingers, deformed limbs, static frames, subtitles, cluttered backgrounds) so you're not starting from a blank negative prompt. ref_image is your subject reference; ref_image_urls is an alternative if your references are already hosted somewhere rather than loaded locally, one URL per line. first_frame and last_frame are separate optional image inputs for pinning the exact start and end of the generated clip, on top of the general ref_image. match_input_num_frames and match_input_frames_per_second (both true by default) auto-derive num_frames (81–241) and frames_per_second from your driving video; turn them off to set those explicitly instead.

Frame-rate and interpolation controls worth knowing about: num_interpolated_frames, temporal_downsample_factor, and enable_auto_downsample let you insert or reduce frames relative to the raw generation - combined with interpolator_model (rife or film, default film) choosing which interpolation algorithm handles the in-between frames, and auto_downsample_min_fps setting the floor enable_auto_downsample won't drop below. This is a lot of surface area for one node - if you're not chasing a specific frame-rate target, leave these at default and only start adjusting once you know exactly what output cadence you need.

Other knobs: preprocess (on by default) runs pose extraction on your driving video automatically; resolution (auto default, plus 240p through 720p) and aspect_ratio (auto default) set output framing; acceleration, video_quality, and video_write_mode are the same speed/quality/file-size trade-offs seen across this pack's other FAL video nodes.

Outputs: video, width, height, fps, url.

Install and setup: search "ComfyUI-utils-nodes" in ComfyUI Manager, or git clone https://github.com/zhangp365/ComfyUI-utils-nodes into custom_nodes, restart. Needs a FAL API key from fal.ai, via api_key or fal_config.yml.

Where this actually breaks:

• With this many optional inputs, it's easy to leave preprocess on while feeding a driving video that's already a pose-extracted control video - double-preprocessing produces motion that doesn't track correctly. Know what your input video actually is before toggling this.
• timeout defaults to 300 seconds; this endpoint does more work per call than a plain i2v request (pose extraction plus generation plus optional interpolation), so it's genuinely more likely to need the higher end of the 1–3000 second range than the simpler nodes in this pack.
• enable_safety_checker defaults to false here too - flip it on deliberately if you need content filtering rather than assuming it's active.
• With num_interpolated_frames, temporal_downsample_factor, and enable_auto_downsample all touching frame count simultaneously, it's easy to end up with an output frame rate you didn't intend. If your final video's motion looks too smooth or too choppy relative to what you expected, this stack of settings - not the base generation - is the first place to look.
• This is a metered call on a relatively expensive endpoint given everything it's doing under the hood. Test on a short driving-video clip before running your full-length footage through it.