Wan 2.1 LoRA · Video Edit

What it is

A video-to-video restyle. You hand it a clip, describe what the whole thing should look like afterwards, and get the same motion and timing back in a new skin. The tooltip is explicit that the source's motion and timing carry over - so this is not a re-generation with a reference, it's a style transfer with temporal structure intact.

That's a genuinely useful primitive. Rotoscope an actor's live-action take into animation without retiming it. Turn a phone video into something that looks shot on film. Push a legitimately-shot clip through a LoRA that changes the wardrobe, the weather, the era. The catch, as always, is that you're trusting the model with everything you didn't mention.

How it works

The incoming VIDEO gets remuxed to MP4 - no re-encode if it's already MP4, so no quality tax - uploaded to PopNow, and the job runs on their GPUs. The result downloads as a fresh VIDEO value you can save or feed into another stage.

Hard limit worth knowing before you queue: uploads cap at 90MiB per file. A 60-second 4K phone clip will blow straight through that, and the node will tell you the input video is larger than the 90 MiB upload limit. Transcode down, trim, or cut the clip into segments first. This is the number one real-world stumble with the video-input nodes in this pack.

LoRAs arrive through the usual route: a PopNow LoRA node per adapter, up to three, source pointing at a local .safetensors, a HF repo id or page, a Civitai page, or an https link.

The inputs that matter

video_url is the source clip, and it's the reason the node exists - motion and timing come from here, not from your prompt.

strength is the dial to learn, and it defaults to a high 0.9 on a 0.1–1 range. That default means "rebuild this from my description". If you want a light grade - keep the faces, keep the room, just change the look - drop to 0.4–0.6 and describe the look rather than the scene. Most first attempts fail because someone leaves 0.9 and writes a prompt about a totally different scene, then wonders why the person changed.

prompt describes the finished clip: subject, wardrobe, setting, style. Since motion is already decided, spend the words on appearance.

resolution is 480p or 720p, duration_seconds 5 or 10, guidance_scale (default 5) and steps (default 30, max 40) behave as they do on the other Wan 2.1 nodes, and flow_shift should be left at -1 to let the model pick 3 at 480p or 5 at 720p. negative_prompt is your "keep this out" field. Outputs are video and task id.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui   # or install "PopNow" from Manager

Restart, drop a key from popnow.ai/api/console/keys into Settings > PopNow (or export POPNOW_API_KEY on a headless box), restart again. Nothing to download. If you're wiring a clip in, remember ComfyUI has to be able to read it - a standard video loader or the pack's own saved output both work, and ComfyUI/output/popnow_<task>.mp4 is where PopNow videos land.

Where it goes wrong

Restyling a clip is billed per request, and 10 seconds at 720p is roughly twice the entry rate of 5 seconds at 480p - so do your strength experiments at 480p/5s on a trimmed clip, then run the full thing once.

If the output looks washed out compared to the source, you pushed strength high enough to rebuild detail the prompt didn't ask for; lower it and describe texture. If motion wobbles or trails, that's the model reinterpreting the source at a high strength, same fix.

LoRA mismatches stay silent: this slot wants Wan 2.1 14B adapters. And a 40201 is an empty wallet, not a bug in your graph.