English · 한국어 · 中文

image-to-video-wan

This service takes an image and a text description and returns a short video made with Wan-AI/Wan2.2-TI2V-5B.

Quick Start

It needs model-compose 0.4.120 or later.

Install with uv:

uv pip install model-compose

Or with pip:

pip install model-compose

Clone this repository:

git clone https://github.com/MindrLabs/image-to-video-wan
cd image-to-video-wan

Run it:

model-compose up

The API listens on http://localhost:8080/api and the Gradio UI opens at http://localhost:8081.

Summary

The RTX 4090 made each 5.04-second video in a median of 526 seconds, meeting the model card's speed claim for a single consumer GPU.

Machine	Status	RTF	Peak memory (MiB)
DGX Spark	Measured	281	30013 (VRAM)
RTX 4090	Measured	104	23942 (VRAM)
MacBook M1 16GB	Not measured	—	—

Setup

Measured with model-compose 0.4.120. Statistics use 95% intervals, and speed differences under 1.05× are not called.

Inputs	Weights	Claim rules
3 inputs × 1 repeat	Wan-AI/Wan2.2-TI2V-5B	fps = 24; pixels = 720p; aspect_ratio_error ≤ 0.05

Results

Image to Video

An image starts the video and a text description of the motion goes in with it.

Claim	Check	Machine	Pass
TI2V-5B generates 24fps video.	Checked by rule	DGX Spark	✓
TI2V-5B generates 24fps video.	Checked by rule	RTX 4090	✓
TI2V-5B generates 720P video.	Checked by rule	DGX Spark	✓
TI2V-5B generates 720P video.	Checked by rule	RTX 4090	✓
The TI2V-5B video keeps the aspect ratio of the input image.	Checked by rule	DGX Spark	✓
The TI2V-5B video keeps the aspect ratio of the input image.	Checked by rule	RTX 4090	✓
TI2V-5B turns an image into a video that follows the image.	Judged by a person	—	✓
TI2V-5B generates a 5-second 720P video in under 9 minutes on a single consumer-grade GPU.	Judged by a person	—	✓

We measured on DGX Spark, RTX 4090. We could not measure on MacBook M1 16GB. RTF is the processing time divided by the length of the audio or video, so lower is faster. Each cell is the median over measured calls, leaving out any call made during other GPU work.

Machine	RTF	Seconds per call	Peak memory (MiB)	aspect_ratio_error	duration_sec	fps	height	pixels	width
DGX Spark	281	1419	30013 (VRAM)	0.0256	5.04	24	896	882688	1013
RTX 4090	104	526	23942 (VRAM)	0.0256	5.04	24	896	882688	1013

A girl in a school uniform stands still before a giant aquarium window, looking up as a huge glowing jellyfish drifts slowly with its long tentacles trailing, small fish swim past, sunlight ripples down through the water, static camera.

A girl in a school uniform drifts down through the air toward a sunlit crosswalk, her short hair and pleated skirt flutter in the wind, she reaches out with both arms, rippling light shimmers across the asphalt below and her shadow grows on the road, the camera slowly arcs a little around her, shifting the viewing angle.

A girl sits on a train bench gazing out the window as a giant turtle glides past over a ruined, overgrown city, its eye blinking slowly, warm sunbeams shift across the carriage as the train sways gently, static camera.

Recommendations

• The RTX 4090 used up to 23942 MiB of GPU memory, so free that much before you start.

Limitations and What We Did Not Measure

• Each of the 3 inputs was generated only once, so the person's judgment of the motion applies only to these outputs.

Measurement Conditions

Item	Detail
Constraint	The measurements set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True in the server environment, because without it the VAE decode ran out of memory on the RTX 4090.
Constraint	Measurements ran with WAN_CPU_OFFLOAD=true, while model-compose.yml defaults it to false.
Claim rules	Claims with a rule were checked on each machine against the number their source gives.

Not Measured

Subject	Item	Reason
MacBook M1 16GB	Every measurement	The model's flash-attn dependency could not be installed on Apple Silicon.
DGX Spark vs RTX 4090	RTF	3 inputs measured on both without other GPU work, at least 20 are needed.
DGX Spark vs RTX 4090	aspect_ratio_error	3 inputs scored on both, at least 20 are needed.

Unchecked Model Card Claims

• The single-GPU TI2V-5B command runs on a GPU with 24GB of VRAM such as the RTX 4090.

License

Covers	License	Commercial use
This service	LICENSE	✓
Wan2.2-TI2V-5B weights	MODEL_LICENSE	✓