Act Two
Video Available Full comparison ↗ runway/act-two · by Runway
· Diffusion transformer (DiT)
Pricing — 1 offering(s)
About this offering: Runway Act Two — a character performance capture and transfer model. Successor to Act One (which ran on Gen-3 Alpha). Accepts a driver performance video (actor's motion/expressions/audio) plus a character reference (image or video) and outputs the character performing the driver's scene. NOT a prompt-driven video editor; the "prompt" is the driver video, not text. API endpoint: POST /v1/character_performance. Output: up to 10s at 1080p, 24fps. Driver video: up to ~10–30s depending on plan tier. Launched 2025-07-21 on the Runway Developer API.
Standard
- $0.05 / second
Current
2025-07-21 → present
5 credits/second × $0.01/credit = $0.05/s. Output billed per second of generated video, not per second of…
5 credits/second × $0.01/credit = $0.05/s. Output billed per second of generated video, not per second of driver input. API endpoint: POST /v1/character_performance. Launched 2025-07-21.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Act Two fits into a cost-aware routing setup
See how →Capability profile
How Act Two rates across core capability dimensions, with the task-level evidence behind each rating.
Act Two is driven by a video (driver performance), not a text prompt. A text prompt is optional and plays a secondary role in the output. The primary creative input is the driver video + character reference. "Prompt adherence" as a dimension does not apply in the standard sense.
Performance transfer model — temporal consistency is fundamental to preserving the actor's motion and expression faithfully across frames. The driver video provides an explicit temporal reference.
Motion capture is the core use case. Transfers gesture, locomotion, and expressions from driver video to character with full-body support (when character input is an image). Runway highlights processing under 30 seconds per clip.
Audio is transferred from the driver performance video to the output. The generated character's lip sync, speech patterns, and delivery are derived from the driver's audio — not synthesized independently.
Camera motion from the driver video is preserved when the character reference is a video; when character reference is an image, the output uses a default camera. No explicit camera parameter API.
Character identity preservation is the entire purpose of this model. The character reference (image or video) anchors the visual identity; Act Two transfers motion onto it while maintaining appearance.
1080p output. No 4K option.
Processing under 30 seconds per clip reported. At $0.05/s, very cost-effective.
Output clips up to 10 seconds. Driver performance video accepted up to ~10–30 seconds depending on plan. No multi-clip chaining mode documented.
When the character reference is an image, Act Two animates it with full-body gesture control from the driver video — substantially more capability than standard I2V start-frame animation.
Not applicable in the standard T2V/I2V sense; no benchmark data.
At $0.05/s — equal to Gen-4 Turbo — this is one of the lowest-cost per-second offerings in the registry despite being a specialized high-value capability.
Act Two does NOT perform prompt-driven modification of an existing video clip. It transfers performance from a driver video onto a character reference — a distinct capability from video editing. Do not conflate character performance transfer with the video-editing dimension (which covers inpainting, style rewriting, element replacement). rating: weak because the model produces a new video rather than modifying an existing one in a targeted way.
Operator guidance
Choose Act Two when you need to animate a character (image or video) using a live performance — motion, expressions, gestures, and audio all transfer from the driver video. At $0.05/s it is extremely cost-effective for this specialized use. This is NOT a general-purpose video editor: Act Two does not modify existing footage via text prompts (see video-editing dimension). Compare with Kling 3.0's element reference support, which keeps a character's identity across a new generation but does NOT transfer motion from a performance driver.
Use cases
- Animating character images with realistic full-body performances from a webcam/actor video
- Transferring professional actor performances to animated or photorealistic characters
- Multi-character dialogue scenes with a single actor driving multiple characters
- Dubbing and re-performance for localization workflows
Limitations
- Driver video required for every generation — no text-only mode
- When character reference is a video, only facial/expression control is available (no full-body gestures)
- When character reference is an image, environment and camera motion come from the driver video
- Output capped at 10 seconds
- No mask-guided inpainting or prompt-driven element replacement
- Not a general video editor — limited to performance transfer use cases