Prefer a visible face and body, simple occlusion and enough space around the subject.
Wan-Animate-2
End-to-end character animation from an image and a driving video
Wan-Animate-2 directly consumes a reference character, motion video and text description inside a redesigned 14B Diffusion Transformer. This page combines the exact model guide with a public Hugging Face ZeroGPU demo.
Reference character
Driving video
Appearance and background prompt
Wan-Animate-2 Animated character video
Create a short Wan-Animate-2 video online
Upload one character image and a short driving video. The embedded Space uses the distilled Wan-Animate-2 Diffusers checkpoint on Hugging Face ZeroGPU; no local installation is required.
A real ZeroGPU model demo
The Space loads Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers. It does not forward the request to DashScope, and its /animate endpoint is also exposed through Gradio.
Loading the Wan-Animate-2 Space…
Uploads leave wan2.video
Your reference image, driving video and prompt are submitted directly to the third-party Hugging Face Space. Do not upload private, confidential or unlicensed material. Queue, retention and moderation are controlled by Hugging Face and the Space owner.
The public demo trims the driving video to 1–5 seconds and converts it to 24fps.
Use the prompt for character clothing, identity cues and background; motion comes from the driving video.
Try 320×480 and 4–6 steps before spending quota on a larger render.
What Wan-Animate-2 changes
Released on August 7, 2026, Wan-Animate-2 is a new character-animation framework rather than a renamed build of the first Wan2.2 Animate. It removes the intermediate motion-extraction stage and puts driving-video conditioning inside the video DiT.
Official architecture: reference appearance, driving motion and text conditioning enter the redesigned video Diffusion Transformer directly.
Direct driving-video conditioning
The model consumes the motion source in the DiT instead of depending on a separate pose or motion extractor.
Identity preservation
Reference-image features are designed to keep character appearance stable while transferring movement.
Text-driven viewpoint control
Text can influence output camera perspective so it is not fully locked to the driving video's viewpoint.
Distilled fast path
A distilled checkpoint reduces the sampling workload; the embedded Space uses that Diffusers variant.
Exact model used by this page
The iframe runs Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers. Older fal.ai, Replicate and community Wan2.2 Animate endpoints may still run the previous-generation model.
From reference image to animated result
The free demo exposes the shortest usable production path while keeping the model's three conditioning inputs visible.
Upload a single image that defines the face, clothing, proportions and initial visual identity.
Upload a 1–5 second driving clip. Its body, head and camera motion guide the animation.
Write an objective character and background description. Avoid duplicating the motion instruction.
Start with modest resolution and steps, then check identity, limbs, occlusion and camera behavior.
Free demo parameters
These are the current controls in the hugging-apps Space, not the complete limits of the downloadable model.
ParameterCurrent rangeWhat it controls
Duration1–5 s
How much of the beginning of the driving video is animated.
Height and width320–640 px
Each dimension is rounded to a multiple of 16 by the Space.
Inference steps4–20
More steps cost more GPU time; the demo defaults to 6.
Guidance scale1.0–6.0
At 1.0 classifier-free guidance is disabled for a faster distilled run.
Sample shift1.0–12.0
Adjusts the flow-matching sampling schedule; the default is 5.0.
Output frame rate24 fps
The input is normalized and the output MP4 is exported at 24fps.
Seed0+
Reproduce or vary the random generation path.
The official repository supports higher production settings, including 720P on its documented 8× A800 configuration. Do not treat the public Demo sliders as model limits.
Free does not mean unlimited
Hugging Face ZeroGPU assigns a daily GPU allowance by account type. This Space requests an xlarge GPU, and larger frames or more steps can consume the allowance quickly.
Anonymous requests also use a shared pool with stricter limits. Queue availability and quota rules can change without notice.
Space performance estimates
320 × 480 · 4 steps≈ 23s
640 × 480 · 8 steps≈ 237s
These are two measurements embedded in the current Space code, not standardized model benchmarks or a guaranteed wait time.
The same Space is a public Gradio API
The Generate button registers api_name="animate". Inspect the live schema with view_api() before sending files because the Space owner can change parameters.
/animate
from gradio_client import Client, handle_file
client = Client(
"hugging-apps/wan2-2-animate-2-14b"
)
client.view_api()
# Then call api_name="/animate"Discover the current schema
Use client.view_api() or the Space OpenAPI document instead of permanently hard-coding an old argument order.
Handle both files
The reference image and driving video must be uploaded through Gradio file handling before inference.
Expect a queued job
Video generation is asynchronous, GPU intensive and subject to ZeroGPU availability and per-user limits.
Wan-Animate-2 is not the old Wan2.2 Animate
Both animate characters, but the new release changes the conditioning architecture and introduces a distinct model family.
DimensionWan2.2 AnimateWan-Animate-2
Generation
First-generation Wan2.2 Animate workflow
New end-to-end Wan-Animate-2 framework
Motion input
Uses intermediate motion representations
Driving video enters the redesigned DiT directly
Camera
More closely follows the source setup
Adds text-driven viewpoint control
This free demo
Not used
Runs the 14B distilled Diffusers checkpoint
What to know before generating
This is an early, compute-heavy public demo. A successful render still depends on input quality, queue state and GPU quota.
Short demo duration
The embedded Space accepts only the first 1–5 seconds of the driving clip.
Character drift is still possible
Occlusions, fast turns, hands and unusual anatomy can cause identity or structure errors.
Public service availability
The Space can sleep, queue, fail, update or disappear independently of wan2.video.
Rights and privacy
You remain responsible for consent, likeness rights, copyright and lawful use of every uploaded asset.
Local deployment is a different scale
The official reference configuration is designed for high-end multi-GPU systems, not a typical browser or consumer laptop.
- Official default configuration: 8× A800 for 720P
- Officially tested: 2× A800 for 480P
- Base and distilled weights, Diffusers and ComfyUI integration are available
Wan-Animate-2 questions
The practical distinctions between the model, this Space and a production API.
Is the Wan-Animate-2 demo really free?+
The embedded public Space currently uses Hugging Face ZeroGPU and does not charge wan2.video users directly. It is limited by daily account quota, queue capacity and the Space's availability.
Do I need a Hugging Face account?+
Anonymous access is currently possible, but it has a smaller two-minute daily quota and a stricter shared pool. Signing in to Hugging Face can provide the account quota listed above.
Does the demo use the exact new model?+
Yes. Its source code names Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers, which belongs to the new Wan-Animate-2 release.
Can I call it as an API?+
Yes, through the public Gradio /animate endpoint while the Space remains available. Treat it as an experimental demo endpoint, not a production service.
Can I generate long or 720P videos in the iframe?+
No. The current public controls limit duration to 1–5 seconds and each dimension to 320–640 pixels. The downloadable model supports other settings with substantially more hardware.
Model and demo references
Technical details were checked against the official model card, paper, Space source and Hugging Face API documentation on August 29, 2026.