Elastic model: Fastest self-serving models. Wan 2.2

Elastic models are the models produced by TheStage AI ANNA: Automated Neural Networks Accelerator. ANNA allows you to control model size, latency and quality with a simple slider movement. For each model, ANNA produces a series of optimized models:

S: The fastest model, with accuracy degradation less than 2%.

Goals of Elastic Models:

Provide the fastest models and service for self-hosting.
Provide flexibility in cost vs quality selection for inference.
Provide clear quality and latency benchmarks.
Provide interface of HF libraries: transformers and diffusers with a single line of code.
Provide models supported on a wide range of hardware, which are pre-compiled and require no JIT.

It's important to note that specific quality degradation can vary from model to model. For instance, with an S model, you can have 0.5% degradation as well.

Prompt: Massive ocean waves violently crashing and shattering against jagged rocky cliffs during an intense storm with lightning flashes

Resolution: 480x480, Number of frames: 81

S	Original
/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F6799fc8e150f5a4014b030ca%2F7Z1gFce9lMkOfFKrc8UUk.mp4%3C%2Fa%3E%3C%2Ftd%3E	/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F6799fc8e150f5a4014b030ca%2Fb-JwMpD8LhbbvUdU2kjFE.mp4%3C%2Fa%3E%3C%2Ftd%3E

Inference

Compiled versions are currently available only for 81-frame generations at 480x480 resolution. Other versions are not yet accessible. Stay tuned for updates!

To infer our models, you just need to replace diffusers import with elastic_models.diffusers:

import torch
from elastic_models.diffusers import WanPipeline
from diffusers.utils import export_to_video

model_name = "Wan-AI/Wan2.2-T2V-A14B-Diffusers"
device = torch.device("cuda")
dtype = torch.bfloat16

pipe = WanPipeline.from_pretrained(
    model_name,
    torch_dtype=dtype,
    mode="S"
)
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
pipe.to(device)

prompt = "A beautiful woman in a red dress dancing"

with torch.no_grad():
    output = pipe(
        prompt=prompt,
        negative_prompt="",
        height=480,
        width=480,
        num_frames=81,
        num_inference_steps=40,
        guidance_scale=3.0,
        guidance_scale_2=2.0,
        generator=torch.Generator("cuda").manual_seed(42),
    )

    video = output.frames[0]
    export_to_video(video, "wan_output.mp4", fps=16)

Installation

System requirements:

GPUs: H100
CPU: AMD, Intel
Python: 3.10-3.12

To work with our models just run these lines in your terminal:

pip install thestage
pip install 'thestage-elastic-models[nvidia]' --extra-index-url https://thestage.jfrog.io/artifactory/api/pypi/pypi-thestage-ai-production/simple
pip install transformers==4.52.3
pip install diffusers==0.35.1

pip install flash_attn==2.7.3 --no-build-isolation
pip uninstall apex
pip install tensorrt==10.11.0.33 opencv-python==4.11.0.86 imageio-ffmpeg==0.6.0

Then go to app.thestage.ai, login and generate API token from your profile page. Set up API token as follows:

thestage config set --api-token <YOUR_API_TOKEN>

Congrats, now you can use accelerated models!

Benchmarks

Benchmarking is one of the most important procedures during model acceleration. We aim to provide clear performance metrics for models using our algorithms.

Quality benchmarks

We used a benchmark VBench (https://github.com/Vchitect/VBench) to evaluate the quality.

Metric	S	Original
Subject Consistency	0.96	0.96
Background Consistency	0.96	0.96
Motion Smoothness	0.98	0.98
Dynamic Degree	0.29	0.29
Aesthetic Quality	0.62	0.62
Imaging Quality	0.68	0.68

Latency benchmarks

Time in seconds of generation for 480x480 resolution, 81 frames.

GPU	S	Original
H100	90	180

Model tree for TheStageAI/Elastic-Wan2.2-T2V-A14B-Diffusers

Base model

Wan-AI/Wan2.2-T2V-A14B-Diffusers

Quantized

(17)

this model

TheStageAI
/

Elastic-Wan2.2-T2V-A14B-Diffusers