DreamActor M2.0 (SIGGRAPH 2026) is ByteDance's motion-transfer model: one reference image plus one driving video animates any character — humans, cartoons, animals — copying motion, expressions, and lip sync. Unlike M1, it's pose-free (no skeleton extraction), enabling multi-character scenes and human-object interaction. Official pricing is $0.05/s; weights are closed and output specs unpublished. It runs inside the invideo agent.
Updated August 2026
DreamActor M2.0's headline advance is that it is pose-free: no skeleton extraction, no rig, no pose-estimation step at all. ByteDance's motion-transfer model takes one image of a character — a person, a cartoon, an animal, a brand mascot — plus one driving video of a performer, and animates the character with that performer's motion, facial expressions, and lip movements. The research behind it (arXiv 2601.21716, January 2026) was accepted to SIGGRAPH 2026. Inside invideo, it shows up as "DreamActor M2.0" in the agent's video tab.
What does DreamActor M2.0 actually do?
The model performs cross-identity motion transfer. You supply a reference image (portrait or full body) that defines who moves, and a driving video that defines how they move. DreamActor M2.0 copies the body motion, facial expressions, and lip sync from the video onto the character in the image while preserving the character's identity and the image's background. ByteDance's BytePlus platform markets it as "the motion control feature driven by the large model" — meaning a video foundation model learns the transfer end to end instead of a hand-engineered animation pipeline.
The lineage matters. DreamActor-M1 (arXiv 2504.01724, April 2025, accepted to ICCV 2025) came from the same ByteDance Intelligent Creation team and animated human photos using hybrid guidance — implicit facial representations combined with 3D head spheres and 3D body skeletons. M2.0, published in January 2026 with academic co-authors from ICT-CAS, UCAS, and Southeast University, throws the explicit 3D scaffolding away.
Why pose-free motion transfer is the technical leap
Skeleton-based animation pipelines share a bottleneck: they first extract a pose sequence from the driving video, then render the character to match it. That works tolerably for standard human bodies and fails for everything else — flowing cloth, animal anatomy, cartoon characters with stylized proportions, and any motion the skeleton representation simply cannot encode.
DreamActor M2.0 reformulates motion conditioning as spatiotemporal in-context learning on raw RGB frames. Per the project page, two ideas carry the system: reference appearance and motion cues are fused in a unified latent space on top of a video foundation model, and the training data problem is solved by self-bootstrapped synthesis of pseudo cross-identity pairs. The paper argues this breaks M1's "see-saw" trade-off between identity preservation and motion fidelity, and it generalizes to non-humanoid characters, multi-character synchronization, and human-object interaction. ByteDance also introduced AWBench, a benchmark for this animate-anything task, alongside the paper.
DreamActor-M1 vs M2.0: what changed
| DreamActor-M1 | DreamActor M2.0 | |
|---|---|---|
| Paper / venue | arXiv 2504.01724 (Apr 2025), ICCV 2025 | arXiv 2601.21716 (Jan 2026), SIGGRAPH 2026 |
| Motion guidance | Hybrid: implicit facial features + 3D head spheres + 3D body skeletons | Pose-free — spatiotemporal in-context learning on raw RGB |
| Subjects | Human photos (portrait to full body) | Any character: humans, cartoons, animals, stylized figures |
| Multi-character | No | Yes, synchronized |
| Human-object interaction | No | Yes |
| Weights | Never released | Not released |
| Commercial access | None | Official API from ~Feb 2026 |
Every M2.0 column entry is a removal — the skeleton, the humans-only limit, the single-character limit — except the last row, where a commercial API finally appears. That is the whole M1-to-M2.0 story in one table.
DreamActor M2.0 specs (and what ByteDance won't say)
| Spec | Detail (as of August 2026) |
|---|---|
| Inputs | 1 reference image + 1 driving video |
| Output resolution | Not published by ByteDance |
| Max output duration | Not published by ByteDance |
| Official pricing | $0.05 per second of output via BytePlus, single-concurrency tier |
| Weights | Closed; no open-source release (M1 was likewise never released) |
| Benchmark | AWBench (ByteDance's own, introduced with the paper) |
Two honesty notes. ByteDance has not published output resolution or maximum clip duration for the commercial endpoints — specific numbers elsewhere are third-party claims. And the only published quality benchmark is ByteDance's own AWBench; independent comparisons against other motion-transfer systems were not available as of August 2026.
What should you use DreamActor M2.0 for?
- Character animation from a single design. One illustration is enough — no rigging. Pair it with a character generator to design the character, then drive it with recorded performance.
- Dance and UGC trend formats. Record the trend once, then map the performance onto any character. This is the classic motion control workflow: performance as an input layer, identity as a swappable one.
- Animating mascots and non-human characters. The pose-free approach is exactly what animals, blobs, and stylized mascots need, feeding a broader AI animation pipeline without a 3D rig ever existing.
How to pick a good driving video
The driving video is the model's entire motion specification, so its quality dominates the result. Use a single, fully visible performer; keep the camera locked so subject motion isn't confused with camera motion; favor broad, readable gestures for stylized characters over subtle micro-movement; and match framing — a portrait reference pairs best with a head-and-shoulders clip, a full-body reference with a full-body performance.
DreamActor M2.0, answered
Is DreamActor M2.0 open source?
No. Neither M1 nor M2.0 has released weights — this is a papers-plus-API model family. Access is through ByteDance's official API and platforms that integrate it.
How much does DreamActor M2.0 cost?
The official BytePlus price is $0.05 per second of generated video on a single-concurrency tier, as of August 2026.
What's the difference between DreamActor M1 and M2.0?
M1 (2025) animated human photos using 3D skeletons and head spheres. M2.0 (2026) drops pose extraction entirely, learns motion from raw video frames, and extends to non-human characters, multiple characters, and human-object interaction.
Can DreamActor animate animals and cartoons?
Yes — that's its defining capability. With no skeleton step, subjects that don't fit a human rig are in scope, per the SIGGRAPH 2026 paper.
What resolution does DreamActor M2.0 output?
ByteDance hasn't published output resolution or duration limits for the commercial endpoints as of August 2026. Treat unofficial numbers with caution.
Where can you use DreamActor M2.0?
Open the video tab in the invideo agent and you'll find it listed as "DreamActor M2.0" — upload your character image and driving clip, and the agent handles the rest, dropping the output straight into a timeline. It sits in the roster on invideo's AI models index, and since invideo is the AI video platform that gives serious creatives every major model in one place, the character image and driving footage can be generated in the same workspace.
Version history: first published August 2026, covering DreamActor-M1 (April 2025, ICCV 2025) and DreamActor M2.0 (January 2026, SIGGRAPH 2026; commercial API from ~February 2026).