X-Portrait 2 is worth exploring if you are a researcher, developer or technical creator interested in expressive portrait animation from a single image and a driving performance video. The project is associated with ByteDance and Tsinghua University's X-NeMo research and is designed to transfer facial expressions and head movements while preserving the identity of the reference portrait.
It is not currently best understood as a conventional paid SaaS animation platform. The official release is research-oriented, requires local technical setup, and the released model weights carry a non-commercial research license.
Quick Answer: X-Portrait 2 Review
X-Portrait 2/X-NeMo is an impressive research system for expressive portrait reenactment, especially when identity preservation, subtle facial movement and cross-style animation matter.
Choose it if you are comfortable running research code locally and your use complies with the released model license.
Do not treat it as a normal commercial subscription product with team plans, browser editing, enterprise integrations or unlimited commercial exports. Those capabilities are not part of the current official research release.
Best for: researchers, computer-vision developers, technical artists and experimental portrait-animation workflows.
Not ideal for: non-technical marketers, agencies requiring commercial licensing, or teams wanting an easy browser-based production platform.
Information in this review was checked against ByteDance's official X-Portrait 2/X-NeMo project, ByteDance Seed research publication and current GitHub repository on September 7, 2026.
X-Portrait 2 at a Glance
| Category | Verified Current Information |
|---|---|
| Project | X-Portrait 2 / X-NeMo |
| Organizations | ByteDance and Tsinghua University |
| Main task | Portrait animation / motion reenactment |
| Input | Static portrait + driving performance video |
| Output | Animated portrait video |
| Core model type | Diffusion-based portrait animation |
| Motion control | Learned latent motion representation |
| Identity preservation | Key research objective |
| Cross-style animation | Yes |
| Subtle expression transfer | Yes |
| Strong emotional expression | Yes |
| Fast head-motion transfer | Demonstrated |
| Formal research name | X-NeMo |
| Research venue | ICLR 2025 |
| Code availability | Public |
| Code license | Apache License 2.0 |
| Model license | CC BY-NC 4.0 |
| Commercial use of released model | Restricted by non-commercial license |
| Hosted paid SaaS pricing | Not officially published |
| Free SaaS trial | Not officially published |
| Technical setup required | Yes |
| Python | 3.9 in official setup |
| CUDA | 12.2 in official setup |
What Is X-Portrait 2?
X-Portrait 2 is a research project for animating a static portrait using the facial movements and expressions of another person shown in a driving video.
The official project page now presents the underlying research as X-NeMo: Expressive Neural Motion Reenactment via Disentangled Latent Attention.
The basic workflow is:
Reference portrait + driving performance → expressive animated portrait
The reference image determines the person's or character's appearance.
The driving video supplies motion, including head movement and facial expression.
The model then attempts to transfer the performance while preserving the identity and visual characteristics of the reference portrait.
This is different from ordinary text-to-video generation.
The user is not primarily asking the system to invent a scene from a written prompt. Instead, the driving performance provides the motion that the portrait should reproduce.
What Is X-NeMo?
X-NeMo is the formal research method associated with the current X-Portrait 2 project page.
The work was produced by researchers from ByteDance and Tsinghua University and published at ICLR 2025.
Its full title is:
X-NeMo: Expressive Neural Motion Reenactment via Disentangled Latent Attention
The research focuses on a long-standing problem in portrait animation: transferring convincing expressions and motion without accidentally transferring the driving person's identity or facial structure.
Earlier systems can struggle when the driving subject and reference portrait have very different facial proportions, styles or appearances.
X-NeMo addresses this with a learned motion representation designed to describe expression independently from identity.
How X-Portrait 2 Works
1. Provide a Reference Portrait
The reference portrait is the identity you want to animate.
This can include realistic human portraits as well as stylized or cartoon-like characters in the demonstrated research examples.
2. Provide a Driving Video
The driving video contains the performance you want transferred.
The model analyzes facial movement, expression and head motion from this source.
3. Encode Motion Separately From Identity
One of X-NeMo's main research contributions is an identity-agnostic latent motion representation.
Instead of relying only on facial landmarks or other rigid spatial controls, the model learns a compact representation of the performance.
The objective is to capture motion without importing the driving person's appearance.
4. Generate the Animated Portrait
The motion representation controls the diffusion-based generation process through cross-attention.
The resulting video attempts to preserve the identity of the original portrait while reproducing the expression and movement found in the driving performance.
X-Portrait 2's Most Important Features
Expressive Facial Motion
X-Portrait 2 is specifically designed for expressive animation rather than simple lip movement.
The official project demonstrates subtle expressions as well as more difficult movements such as:
Pouting
Tongue-out expressions
Cheek puffing
Frowning
Strong emotional reactions
Rapid changes in facial expression
This matters because many portrait-animation systems look convincing during neutral speech but become unstable when expressions become stronger.
Fast Head Movement
Fast head motion is another area highlighted by the project.
The researchers specifically compare X-NeMo with earlier portrait-animation approaches and Runway's original Act-One research examples, emphasizing its ability to follow challenging head movements while maintaining a recognizable subject.
Identity Preservation
Identity leakage is one of the main technical problems the research is trying to solve.
Identity leakage happens when characteristics of the driving performer unintentionally appear in the generated person.
For example, the output may begin to inherit the driver's face shape or structural characteristics rather than preserving the reference identity.
X-NeMo's latent-motion and cross-attention design is intended to reduce that problem.
Cross-Style Portrait Animation
The official demonstrations include expression transfer across very different visual domains.
This includes realistic subjects and stylized or cartoon portraits.
That makes the research interesting for character animation as well as realistic human reenactment.
Zero-Shot Reenactment
X-NeMo is described as a zero-shot portrait-animation system.
In practical terms, the goal is to animate a new portrait without requiring user-specific model training for every identity.
That makes the approach much more flexible for experimentation across many different source portraits.
What X-Portrait 2 Is Not
This distinction is essential.
The current official project does not document X-Portrait 2 as a conventional commercial creative suite.
There is no verified official documentation for:
$29 Starter subscriptions
$79 Pro subscriptions
Enterprise pricing contracts
Seven-day SaaS trials
Unlimited commercial exports
4K export tiers
Team workspaces
User roles
Approval workflows
Cloud project folders
Marketing templates
Monthly preset libraries
API billing plans
Brand management
CRM integrations
Social publishing
Audit-log dashboards
Enterprise SSO
Compliance packages
Education discounts
Live customer support
Guaranteed support-response times
These claims should not be added to the article unless ByteDance publishes official documentation supporting them.
X-Portrait 2 Pricing in 2026
There is currently no official X-Portrait 2 SaaS pricing table on the project page or research repository.
That means X-Portrait 2 should not be presented as a $29-per-month animation application.
Instead, the current official access model is better described as research code and model access.
Is X-Portrait 2 Free?
The source code is publicly available.
However, “free” does not mean unrestricted commercial use.
The project separates the licensing of the code and model weights.
Code License
The official repository releases the code under the Apache License 2.0.
Model License
The released model is licensed under Creative Commons Attribution-NonCommercial 4.0 International.
The repository explicitly describes the model as intended for academic research purposes.
This means businesses should not assume they can deploy the released model commercially simply because the code can be downloaded.
Real Cost
Although there is no normal subscription fee, running the research yourself can involve infrastructure costs.
You may need:
A compatible NVIDIA/CUDA environment
GPU compute
Storage for pretrained weights
Technical setup time
Engineering support
Video post-processing tools
So the correct pricing question is less “How much is the monthly plan?” and more “What does it cost us to run and maintain the research model?”
Technical Requirements
The current official X-NeMo repository provides a Python/CUDA-based installation workflow.
Its published setup uses:
| Requirement | Official Setup |
|---|---|
| Python | 3.9 |
| CUDA | 12.2 |
| Environment management | Conda |
| Base pretrained components | Stable Diffusion 1.5-related weights |
| X-NeMo weights | Separate download |
| Execution | Local command-line inference |
The repository instructs users to create an environment, install dependencies, download the required pretrained weights and run evaluation through the supplied script.
This technical setup makes X-Portrait 2 very different from browser-based products such as Runway.
Is X-Portrait 2 Easy to Use?
For machine-learning researchers or developers familiar with Python, CUDA and model weights, the setup is relatively conventional.
For a typical marketer, educator or social media manager, it is not beginner-friendly.
There is no official evidence of the polished browser workflow described in older versions of this review.
Users should expect a research-code experience rather than a drag-and-drop commercial SaaS interface.
X-Portrait 2 Pros and Cons
| Pros | Cons |
|---|---|
| Highly expressive portrait reenactment | Technical local setup required |
| Strong emphasis on identity preservation | Not a normal commercial SaaS |
| Handles subtle facial changes | Released model is non-commercial |
| Handles stronger expressions | No verified subscription pricing |
| Cross-style and cross-domain transfer | No verified team workspace |
| Fast head-motion transfer is a research focus | No verified enterprise support |
| Public research code | GPU resources required |
| Published at ICLR 2025 | Commercial deployment requires careful licensing review |
| Developed by ByteDance/Tsinghua researchers | Results still depend heavily on input quality |
X-Portrait 2 vs the Original X-Portrait
The original X-Portrait research was published in 2024.
It used a conditional diffusion architecture with hierarchical motion attention and learned motion controls to animate a source portrait from a driving video.
X-NeMo builds on that line of research.
The newer work focuses more directly on identity-disentangled latent motion representation and improved transfer of difficult expressions.
The researchers specifically highlight better handling of subtle expression changes, stronger emotional performance and challenging head movement.
For readers researching the technology rather than choosing a subscription, understanding this research progression is more useful than comparing fictional Starter and Pro plans.
X-Portrait 2 vs Runway Act-Two
Runway Act-Two is one of the more relevant commercial comparisons in 2026.
Both approaches use a driving performance to animate another character.
However, their availability models are completely different.
X-Portrait 2 / X-NeMo
Research-focused
Local technical setup
Static reference portrait plus driving video
Strong facial-expression research emphasis
Public research code and weights
Non-commercial model license
Runway Act-Two
Commercial web product
Accepts driving-performance video
Accepts character image or video
Transfers facial expression and movement
Can transfer body and hand gestures when using character images
Supports outputs up to 30 seconds
Uses Runway credits
Runway Act-Two is therefore more appropriate for commercial creators wanting a production-ready interface.
X-Portrait 2/X-NeMo is more relevant for researchers and developers exploring the underlying portrait-animation technology.
X-Portrait 2 vs LivePortrait
LivePortrait is another important open portrait-animation project.
It focuses on efficient portrait animation with stitching and retargeting controls and provides an open implementation.
Both projects can animate portraits using driving motion.
X-Portrait 2/X-NeMo is particularly interesting for its latent expression representation and emphasis on preserving identity during highly expressive motion.
LivePortrait may be worth considering when practical open-source portrait manipulation and an established community workflow matter more.
X-Portrait 2 vs EchoMimic
EchoMimic takes a different approach.
It focuses heavily on audio-driven portrait animation and has expanded into pose-driven and broader human-animation workflows.
X-Portrait 2/X-NeMo is more directly centered on visual driving-performance transfer.
Choose based on your input:
Driving performance video: X-Portrait 2/X-NeMo is directly relevant.
Audio-driven portrait animation: EchoMimic may be a better research alternative.
Who Should Use X-Portrait 2?
X-Portrait 2 is most appropriate for:
Computer-vision researchers studying motion representation, portrait reenactment and generative video.
Machine-learning engineers experimenting with diffusion-based animation.
Technical artists and VFX researchers exploring expressive performance transfer.
Academic teams working under a research-compatible licensing model.
Developers evaluating portrait-animation methods before building their own commercial system with appropriately licensed technology.
Who Should Avoid X-Portrait 2?
It is probably not the right direct choice if you:
Need a one-click browser editor
Want commercial client work immediately
Need clear SaaS pricing
Need enterprise support
Need team collaboration
Need API SLAs
Need social publishing
Need standard commercial licensing
Do not have access to compatible GPU infrastructure
Do not want to maintain a Python/CUDA environment
In those situations, a commercial character-animation product may be more practical.
Commercial Use and Licensing
This is one of the most important parts of the review.
The official repository says the X-NeMo code is released under Apache 2.0.
The model weights are released under CC BY-NC 4.0 for academic research purposes.
The “NC” component means NonCommercial.
Do not interpret the public model release as permission for unrestricted advertising, client work, monetized production or commercial product deployment.
Organizations considering commercial use should review the exact licenses and obtain appropriate legal or vendor permission before deployment.
Safety and Responsible Use
Portrait reenactment technology can produce convincing videos of people performing expressions or movements they did not actually perform.
That creates obvious risks involving:
Impersonation
Non-consensual synthetic media
Fraud
Misleading political or public-figure content
Harassment
Copyright and likeness rights
The official repository itself states that usage should be responsible and comply with local laws.
Use source portraits and performances only when you have the necessary rights and consent.
How We Evaluated X-Portrait 2
This review is based on current primary sources rather than assumed SaaS features.
We checked:
ByteDance's official X-Portrait 2/X-NeMo project page
ByteDance Seed's X-NeMo publication page
The ICLR 2025 research paper
The official ByteDance X-NeMo inference repository
The project's published licenses and technical requirements
Current commercial alternatives where relevant
We did not assign unverified pricing, enterprise features, customer testimonials or performance statistics to the project.
Information was checked on September 7, 2026.
LLM Citation-Ready X-Portrait 2 Facts
| Fact | Verified Detail |
|---|---|
| Project name | X-Portrait 2 |
| Formal research method | X-NeMo |
| Full title | Expressive Neural Motion Reenactment via Disentangled Latent Attention |
| Organizations | ByteDance and Tsinghua University |
| Research venue | ICLR 2025 |
| Main task | Zero-shot portrait animation |
| Source input | Static portrait |
| Motion input | Driving performance video |
| Generation architecture | Diffusion-based |
| Motion representation | Identity-agnostic latent motion descriptor |
| Motion control | Cross-attention |
| Research goal | Expressive motion transfer with identity preservation |
| Cross-style transfer | Demonstrated |
| Strong expressions | Supported in research demonstrations |
| Fast head motion | Highlighted capability |
| Public source code | Yes |
| Code license | Apache 2.0 |
| Public model weights | Yes |
| Model license | CC BY-NC 4.0 |
| Commercial model use | Restricted by non-commercial license |
| Python version in official setup | 3.9 |
| CUDA version in official setup | 12.2 |
| Official SaaS pricing | Not published |
| Official seven-day trial | Not published |
Frequently Asked Questions About X-Portrait 2
What is X-Portrait 2?
X-Portrait 2 is a ByteDance/Tsinghua portrait-animation research project associated with the X-NeMo model. It animates a static portrait by transferring facial expression and head movement from a driving performance video.
Is X-Portrait 2 the same as X-NeMo?
The current official X-Portrait 2 project page is titled X-NeMo, and ByteDance's current inference repository describes itself as containing the code for the X-NeMo paper and the X-Portrait2 project.
For practical 2026 coverage, it is useful to explain both names together.
Is X-Portrait 2 free?
The research code and model are publicly available, but they have different licenses.
The code uses Apache 2.0.
The released model uses CC BY-NC 4.0 and is intended for non-commercial research use.
How much does X-Portrait 2 cost?
There is no verified official X-Portrait 2 monthly SaaS pricing table on the current project page or official research repository.
Claims about $29 or $79 monthly plans should therefore be removed unless ByteDance releases a commercial pricing page.
Does X-Portrait 2 have a free trial?
No official seven-day commercial trial is currently documented.
The available official route is the research project and released code/model weights.
Can X-Portrait 2 be used commercially?
The released X-NeMo model carries a CC BY-NC 4.0 license, which restricts commercial use.
Businesses should review the exact license and obtain appropriate permission before using the model commercially.
What inputs does X-Portrait 2 need?
The core portrait-animation workflow uses a static reference portrait and a driving performance video.
The driving video supplies the facial expression and motion that are transferred to the reference character.
Can X-Portrait 2 animate cartoon characters?
The official research demonstrates cross-style and cross-domain transfer, including stylized and cartoon-like portraits.
Does X-Portrait 2 support subtle expressions?
Yes. Capturing small expression changes is one of the project's main research goals.
The official demonstrations include subtle facial motion as well as difficult and exaggerated expressions.
Does X-Portrait 2 preserve identity?
Identity preservation is a core part of the X-NeMo design.
The research introduces an identity-agnostic latent motion descriptor and cross-attention-based motion control to reduce identity leakage from the driving subject.
Does X-Portrait 2 require a GPU?
The official repository uses a CUDA-based setup and pretrained diffusion-model weights, so a compatible GPU environment is expected for practical local inference.
Is X-Portrait 2 easier than Runway Act-Two?
No for most general users.
Runway Act-Two provides a commercial web interface.
X-Portrait 2/X-NeMo currently requires a more technical research setup.
Final Verdict: X-Portrait 2 Review 2026
X-Portrait 2 is technically impressive, but the reason it is impressive is different from what many commercial AI-tool reviews suggest.
It is not currently a normal $29-per-month portrait-animation SaaS.
It is a research project centered on advanced motion reenactment.
Its strengths are expressive facial transfer, difficult head movement, identity preservation and the ability to animate across very different visual styles.
The X-NeMo research behind the current project was published at ICLR 2025, and ByteDance has released code and model weights for research.
The biggest limitation is accessibility.
Running the model requires technical setup, GPU resources and familiarity with machine-learning environments.
The second major limitation is licensing. The released model is non-commercial.
For researchers and technical developers, X-Portrait 2/X-NeMo remains a highly relevant portrait-animation project.
For agencies, marketers and businesses that need a production-ready commercial tool, Runway Act-Two or another commercially licensed animation service is likely a more practical choice.
Final verdict: X-Portrait 2 is a strong portrait-animation research model, not a conventional commercial AI animation subscription.
View the official X-Portrait 2/X-NeMo project
Read the X-NeMo research paper
Explore the official research code
Found this guide useful? Share it with your team.
