Alibaba's Character X Generates Synthetic Faces From Scratch, No Photo Needed

Most AI character tools start with a photo of a real person. You upload a face, and the model works from there. Character X, the newest feature shipping inside Alibaba's Wan creative platform, flips that assumption: it generates a completely original face from nothing, no reference photo required.

What Character X actually does

The pitch is simple. Instead of sourcing or uploading a portrait, you describe or configure what you want, and Character X produces a unique, synthetic face that belongs to no real person. Wan is an AI creative platform that aims to lower the barrier to creative work, offering features spanning text-to-image, image-to-image, text-to-video, image-to-video, and image editing. Character X adds a new entry point to that pipeline: original identity generation.

The feature is live now at the Wan Lab Playground. It is positioned for three core use cases:

• Custom avatars for personal or professional use
• Brand characters and mascots that need a consistent, ownable face
• Fictional characters for games, short-form video, or animation pipelines

The face problem Wan has been solving for months

Character X does not appear in a vacuum. Alibaba's Tongyi Lab has been systematically attacking the hardest problems in AI character work throughout the Wan series. Wan's image generation has enhanced its face customization to support all-round control from bone structure and eyes to facial features , including face shape (oval, round, square, rectangular) and eye type (almond, deep-set, round, phoenix) , achieving what the team calls "thousands of faces for thousands of people."

This is particularly useful for creating diverse character groups, virtual idols, or personalized character assets without repetitive facial features. Character X takes that further by making the face-creation step its own dedicated, zero-friction workflow rather than a sub-option buried in image generation settings.

The broader Wan character stack is now substantial:

• Wan2.2-animate , brings animation to a provided character image and a supporting reference video, accurately capturing facial expressions and body movements of the reference character and applying them to a new video; it can also replace a character in a video with one from a source image while preserving their expressions and movements.
• Wan2.2-S2V , converts portrait photos into film-quality avatars capable of speaking, singing, and performing, generating high-quality animated videos from a single image and an audio clip.
• Wan2.6-R2V , enables users to upload a character reference video with both appearance and voice, utilizing text prompts to generate vivid new scenes starring that same character, including multiple subjects together while preserving the distinctive look and sound of the original reference.
• Character X , generates a brand-new face from scratch, no source material needed.

How the underlying tech works

Character X builds on the face generation architecture Alibaba has been refining across the Wan series. During training, the system leverages skeletal information to locate and crop facial regions, and uses self-supervised learning that disentangles identity information from expression information , a critical step to prevent the model from leaking real identities into generated faces. One strategy spatially compresses the facial image into a 1D latent, which reduces the storage of low-level, identity-specific information.

For the animation side of the stack that Character X feeds into, spatially-aligned skeleton signals replicate body motion and implicit facial features extracted from source images reenact expressions, enabling character videos with high controllability and expressiveness; an auxiliary Relighting LoRA preserves the character's appearance consistency while applying the appropriate environmental lighting and color tone.

The workflow: generate a face, then animate it

The real value of Character X is not the face image itself , it is what you do with it next. Once you have a generated face, the full Wan pipeline opens up:

1. Generate an original face with Character X at the Wan Playground
2. Feed that face into Wan2.2-animate to drive it with motion from a reference video
3. Or pass it to Wan2.2-S2V to make it speak or sing from an audio clip
4. Or use Wan2.6-R2V to place it into entirely new scenes via text prompt

The technology can interpret and reproduce subtle movements and expressions that typically require extensive manual effort by skilled animators, speeding up the pace at which realistic-looking characters can be created. Starting with a synthetic face means the entire pipeline is free of any real person's likeness , a meaningful advantage for commercial and brand work.

Where it fits and what it still can't do

Character X is strongest when you need a face that is legally clean, visually unique, and immediately usable across Wan's animation tools. Most tools still struggle with one fundamental problem: keeping a face consistent across frames. Every time you try to generate a branded spokesperson, a recurring character for a YouTube series, or a digital actor for a film pre-visualization, you get a different person in every clip. Character X, paired with Wan's reference-to-video models, is a direct answer to that problem.

The current limitations are worth noting. A clear video with smooth, moderate movements works best downstream; high-energy, erratic motions can sometimes result in blurriness, and the quality of the input directly impacts the quality of the final animation. The feature is also a no-code, playground-first experience , developers who want programmatic access will need to route through Alibaba Cloud's Model Studio API, where the broader Wan character tooling is exposed.

Availability and cost

Character X is available now directly at create.wan.video/lab/playground/character. The Wan series has generated over 30 million downloads on open-source communities and third-party platforms , and the playground has historically offered free credits to new users. Pricing for extended use follows Wan's standard credit model. Wan 2.2 Animate was released as an open-source model under the Apache 2.0 license, with code and weights publicly available so you can run it locally or integrate it into custom workflows without paying for the model itself.

The broader implication is straightforward: the bottleneck for synthetic character pipelines used to be sourcing a face. That step is now a single click.