CJHauser/anima-preview3-onnx

Anima Preview3 ONNX Export

ONNX versions of the Anima Preview3 model components.

Files

• unet/model.onnx - Main diffusion model (2B params, FP32)
• text_encoder/model.onnx - Qwen 0.6B text encoder
• vae/decoder.onnx - Qwen Image VAE decoder

Usage with ONNXRuntime

import onnxruntime as rt

# Load models
unet_session = rt.InferenceSession("unet/model.onnx", providers=["CPUExecutionProvider"])
text_enc_session = rt.InferenceSession("text_encoder/model.onnx", providers=["CPUExecutionProvider"])
vae_session = rt.InferenceSession("vae/decoder.onnx", providers=["CPUExecutionProvider"])

# Run inference
output = unet_session.run(None, {
    "sample": latent_input,
    "timestep": timestep_array,
    "encoder_hidden_states": text_embeddings
})

Notes

• Exported with opset_version=14 for broad compatibility
• Float32 precision for CPU inference
• Latent space assumes 768px resolution (96x96 latents)