Kling 2.6 · Character Animation

This is the motion-transfer node: you give it a still photo of a person and a short clip of somebody moving, and it gives you your person doing that motion. Kling 2.6 Character Animation is the cheaper of the two character-animation nodes in this pack and the one I'd try first, because at 10.35 credits per second you can afford to run it four times and pick the good one.

How it works

The node isn't a local pose estimator plus a local diffusion model. It's a thin client: your photo and your driving clip get uploaded to PopNow, Kuaishou's model does the work, and the finished MP4 is downloaded back into the graph as a VIDEO. No GPU, no ControlNet stack, no OpenPose rig, no ComfyUI dance workflow to assemble. That's the whole trade - you skip the setup and pay per second instead.

Because the model emits video, not latents, the two media inputs are real sockets: image_url takes a ComfyUI IMAGE, video_url takes a VIDEO. That second one is the part beginners trip on. A VIDEO is not a folder of frames - in core ComfyUI you get one from a video loader node (or from another PopNow node's video output), not from Load Image. If you're starting from a file on disk, load it as video, don't unpack frames.

The inputs that matter

image_url - one clear, full-body or half-body picture of the person. JPEG or PNG. Front-facing, evenly lit, no cropped-off feet if the motion is a dance.

video_url - the driving clip, 3 to 30 seconds, and the result is as long as this clip. That single sentence is the cost model: a 30-second driver is roughly 310 credits, a 5-second one about 52. It's also the quality model - you're copying moves, so a clean, well-lit driver with one person in frame beats a chaotic one.

resolution - 720p or 1080p. Higher tiers cost more per second. Start at 720p while you're testing whether the motion reads at all.

prompt - optional notes on the result, not the motion: the setting, the look, what the driving clip doesn't show. Writing a full scene description here fights the driver clip.

Outputs

video (VIDEO) and task id (STRING). Send video to Save Video, or straight into another PopNow video node if you're doing a second pass. The task id is for the API console and for support tickets.

Install

Manager → search PopNow → Install → restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Grab a key from popnow.ai/api/console/keys, put it in Settings > PopNow (or POPNOW_API_KEY), restart again so the node list syncs to your catalog. No model downloads, no dependencies beyond what ComfyUI already ships. Delete the old single-file custom_nodes/popnow_api.py if you had it.

Troubleshooting

Watch the upload limits, because they bite hardest on video inputs. Images go up as PNG, falling back to JPEG if they'd exceed 10 MiB. Video is re-muxed to MP4 and capped at 90 MiB - a 30-second 1080p driver can sail past that, and you get a blunt "larger than the 90 MiB upload limit". Cut the clip or drop the resolution.

Uploads are cached by content hash for 12 hours per API key, so re-running with a different prompt doesn't re-upload the same photo and clip. Good news for iteration, and the reason your second attempt feels faster than the first.

Failures you're likely to see: PopNow 40004 means the model can't serve that combination of settings (usually resolution plus clip length), content_rejected means the model declined - and failed runs are refunded, so retrying with a tamer driver clip costs you nothing but time. A cancelled run is the opposite: it keeps going on PopNow, gets charged if it succeeds, and shows up later in the API console.

One practical note: identity drift is the real failure mode with any character animation, not an error message. Hands, jewellery and profile turns are where it goes. Shoot the source photo close to the angle the motion mostly holds, and keep the driver under 10 seconds for the first test.