Lance Video Edit

Video editing is the capability people most assumed would stay API-only, which is why "Wow, all Edit models are welcomed" was the top reaction to ByteDance's Lance release. Lance Video Edit is that capability in a ComfyUI node: feed it a clip and a sentence - "make the cat look at the camera," "turn day to dusk" - and it rewrites the video accordingly, locally, Apache 2.0, no subscription.

Mechanically it mirrors Lance Image Edit but in the temporal domain. The node takes your frame batch, writes it to a temp MP4 (via imageio, libx264 at 12fps), then builds the official video_edit interleave format - [prompt, video, video], where the last element is the target the model regenerates as the edit. Then it's the same validate_on_fixed_batch path, decoding back to an IMAGE frame stack you can hand to Save Video. The temp-file round-trip is invisible to you but it's why the node feels a beat slower than pure generation: every run encodes a video, decodes the result.

The inputs that matter

• frames - an IMAGE batch from ComfyUI's native Load Video or VideoHelperSuite. Long clips get subsampled to max_frames.
• prompt - describe the change. Required; the node errors if you leave it blank.
• max_frames - default 121, range 5–121, and it steps in 4s because the Wan VAE's temporal stride caps valid counts at 4n+1 (1, 5, 9… 121). It only kicks in when your clip is longer than the cap, and it subsamples uniformly. Lower it to 81 or 41 to trade temporal detail for a much faster, lighter run - the tooltip's advice, and it's good advice when VRAM or patience is tight.
• num_timesteps (30), timestep_shift (3.5), cfg_text_scale (4.0), seed - the same official inference defaults as the rest of the pack.

Wiring and variant

[Lance Model Loader (lance_3b_video)] → [Load Video] → [Lance Video Edit] → [Save Video]

This task requires the video variant (lance_3b_video) in the loader - the image variant won't do temporal editing. Output frames go straight to Save Video for an MP4.

Gotchas

The 40GB VRAM requirement is the headline again; video editing is the heaviest thing this pack does, so budget time accordingly. Your input resolution is also your output resolution, padded to VAE stride - don't feed it a 1080p clip and expect to survive a run on a 24GB card. And remember the 4n+1 frame rule: if your edit result comes back one frame shorter than you fed in, that's the Wan temporal stride doing its thing, not a bug.