Scail 2 on fal: Motion Transfer Without a Skeleton, at $0.10 per Second (October 2026)
Scail 2 copies movement from a driving video onto a character in either animation or replacement mode, at 512p or 704p and 16 frames per second. Why it beat Kling Pro in one of our blind tests, and the limits that keep it from being our default.
For dance and movement copying, Kling Motion Control has been the default for months, and Scail 2 is the challenger with a reason to be taken seriously: in a blind pair we ran in August, we preferred its output to Kling Pro's on that clip, and it is listed at $0.10 per second against Kling's $0.168. That is 40 percent cheaper. Why is it not already our default? Because of four limits, all in the endpoint's schema, that matter more for production than for a single comparison.
What it does
Scail 2 takes a driving video and a character image and makes the character perform the driving video's movement. It has switches that change the job:
| Setting | Options |
|---|---|
mode | animation (animate the character) or replacement (replace the performer in the video) |
| driving method | end_to_end or pose |
subject | human or animal |
| resolution | 512p (512 by 896) or 704p (704 by 1280) |
The pose option drives the character from an extracted pose; end_to_end lets the model work directly from the video. The animal subject is unusual in this category: it lets a pet or a creature take the role of the dancer.
What it costs
| Output length | $0.10 per second | Kling Pro ($0.168) |
|---|---|---|
| 5 seconds | $0.50 | $0.84 |
| 8 seconds | $0.80 | $1.34 |
| 10 seconds | $1.00 | $1.68 |
To reach 1080p you need an upscale pass: Topaz Video Upscale Precision lists $0.01 per second, so a ten-second clip is about $1.10 all in at 1080p, still well under Kling Pro.
The four limits
- Resolution. 704 by 1280 is below 1080 by 1920, so every clip needs upscaling, which is an extra step and an extra place for softness to appear.
- Frame rate. Output is 16 frames per second, noticeably choppier than the 24 or 30 viewers are used to, particularly on fast movement. Frame interpolation can help; it is another step.
- Truncation. Long driving clips are cut short. If the dance you want is 25 seconds, you may get less.
- Audio. We re-attach the trending track ourselves after every motion-transfer run, whichever model produced it.
Together these explain the routing. For a five to eight second hero clip where the blind comparison favoured Scail, it is cheaper and may look better. For a standard-length Reel it needs more post-work than Kling.
How it compares
| Scail 2 | Kling 3.0 Pro Motion Control | |
|---|---|---|
| Price per second | $0.10 | $0.168 |
| Max reference | truncates long clips | up to 30 s in video orientation |
| Output | 512p / 704p, 16 fps | higher resolution |
| Our blind test | preferred on one clip in August | default since April |
| Identity setting | none to configure | character_orientation decides identity |
That last row is a practical difference: with Kling you can get the wrong result by choosing image orientation, as explained in the Motion Control guide. Scail has fewer ways to misconfigure.
Replacement mode versus a video-edit model
replacement mode replaces the performer in the driving video with your character, which overlaps with what video-edit models do. The difference is that Scail follows the pose of the performer and regenerates the body, so the cuts and background are not preserved the way a true editor (Recast, Kling O3 edit) preserves them. Motion transfer versus Recast explains when each wins.
Our status
Scail 2 is active in our catalogue and a candidate for the primary dance model. We have not switched because of the four limits above and because one blind comparison is not a verdict. Our next steps are a ten-second test at 704p with an upscale, and a check of how it handles a clip that is cut short. Until those are done, we describe it as promising and leave Kling in front.
How we would run the test that decides it
Take one ten-second dance clip that Kling Pro already handled, and one character photo with the full body in frame. Run Scail 2 at 704p in animation mode, then upscale the result to 1080p, then put the three outputs (Kling, Scail raw, Scail upscaled) in a row on a phone and watch them with sound. Check frame-rate smoothness on the fastest move, edge softness after the upscale, and whether the clip was truncated. The whole comparison costs under four dollars, and it replaces any opinion, ours included.
Common questions
What is replacement mode?
It replaces the performer in the driving video with your character, rather than animating your character in a new frame.
Does it do animals?
Yes, the subject setting accepts animal as well as human.
Why does the output look choppy?
It runs at 16 frames per second, below the 24 or 30 viewers expect. Frame interpolation can smooth it, as an extra step.
Can it do a 1080p clip?
Not natively. Output is 512p or 704p, so reaching 1080p needs an upscale such as Topaz at $0.01 per second.
Is it your default?
No. Kling Motion Control remains the default while Scail 2's resolution, frame rate and truncation limits are weighed against its price.
How to do this without code: The Dance copy tool routes to the motion-transfer model, and Scail 2 is a candidate we are testing for it. In AutoHustle this is the Dance copy tool: upload the reference clip and one full-body photo (3 credits per second with a 24-credit minimum, quoted before you press Generate). See How to make an AI dance video and AI dance video generator.
Related reading: Wan Motion vs Wan Animate: Two Wan Motion-Transfer Endpoints, One Rejected and One Priced at $0.06 (October 2026) · How AI Dance Face-Swap Videos Are Actually Made
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| Scail 2 | fal-ai/scail-2 |
| Kling 3.0 Pro Motion Control (comparison) | fal-ai/kling-video/v3/pro/motion-control |
Parameters: mode (animation / replacement), driving method (end_to_end / pose), subject (human / animal), resolution 512p (512×896) or 704p (704×1280), 16 fps output. $0.10 per second, read on 18 October 2026.