Function motionctrl_sample

main/evaluation/motionctrl_inference.py:123–190

(
        model, 
        prompts, 
        noise_shape,
        camera_poses=None, 
        trajs=None,
        n_samples=1,
        unconditional_guidance_scale=1.0,
        unconditional_guidance_scale_temporal=None,
        ddim_steps=50,
        ddim_eta=1.,
        **kwargs)

        torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'})

def motionctrl_sample(
        model, 
        prompts, 
        noise_shape,
        camera_poses=None, 
        trajs=None,
        n_samples=1,
        unconditional_guidance_scale=1.0,
        unconditional_guidance_scale_temporal=None,
        ddim_steps=50,
        ddim_eta=1.,
        **kwargs):
    
    ddim_sampler = DDIMSampler(model)
    batch_size = noise_shape[0]
    ## get condition embeddings (support single prompt only)
    if isinstance(prompts, str):
        prompts = [prompts]

    for i in range(len(prompts)):
        prompts[i] = f'{prompts[i]}, {post_prompt}'

    cond = model.get_learned_conditioning(prompts)
    if camera_poses is not None:
        RT = camera_poses[..., None]
    else:
        RT = None

    if trajs is not None:
        traj_features = model.get_traj_features(trajs)
    else:
        traj_features = None

    if unconditional_guidance_scale != 1.0:
        # prompts = batch_size * [""]
        prompts = batch_size * [DEFAULT_NEGATIVE_PROMPT]
        uc = model.get_learned_conditioning(prompts)
        if traj_features is not None:
            un_motion = model.get_traj_features(torch.zeros_like(trajs))
        else:
            un_motion = None
        uc = {"features_adapter": un_motion, "uc": uc}
    else:
        uc = None

    batch_variants = []
    for _ in range(n_samples):
        if ddim_sampler is not None:
            samples, _ = ddim_sampler.sample(S=ddim_steps,
                                            conditioning=cond,
                                            batch_size=noise_shape[0],
                                            shape=noise_shape[1:],
                                            verbose=False,
                                            unconditional_guidance_scale=unconditional_guidance_scale,
                                            unconditional_conditioning=uc,
                                            eta=ddim_eta,
                                            temporal_length=noise_shape[2],
                                            conditional_guidance_scale_temporal=unconditional_guidance_scale_temporal,