add callable object to convert frame into control_frame to reduce cpu memory usage.

[chaowenguo]

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@hlky

[hlky]

Hi @chaowenguo. Could you share a usage example and some benchmarks/profiling to show the reduction in cpu memory usage?

[chaowenguo]

@hlky
example:

import torch, torch_xla, cv2, PIL.Image, diffusers, imageio, builtins, easy_dwpose, rerender_a_video, numpy

strength = [0.9, 0.95, 0.98, 1]

def process(index):
    openpose = easy_dwpose.DWposeDetector()
    with imageio.get_reader('pose0.mp4') as reader, imageio.get_writer(f'out{index}.mp4', fps=reader.get_meta_data().get('fps')) as writer:
        frames = [PIL.Image.fromarray(reader.get_data(_)) for _ in builtins.range(reader.count_frames())]
        controlnet = diffusers.ControlNetModel.from_pretrained('chaowenguo/control_v11p_sd15_openpose', torch_dtype=torch.bfloat16, variant='fp16', use_safetensors=True)
        pipeline = rerender_a_video.RerenderAVideoPipeline.from_single_file('https://huggingface.co/chaowenguo/pal/blob/main/chilloutMix-Ni.safetensors', config='chaowenguo/stable-diffusion-v1-5', safety_checker=None, controlnet=controlnet, use_safetensors=True, torch_dtype=torch.bfloat16, device=torch_xla.core.xla_model.xla_device())
        pipeline.vae = diffusers.AutoencoderKL.from_single_file('https://huggingface.co/chaowenguo/pal/blob/main/vae-ft-mse-840000-ema-pruned.safetensors', torch_dtype=torch.bfloat16, use_safetensors=True).to(torch_xla.core.xla_model.xla_device())
        pipeline.scheduler = diffusers.DDIMScheduler.from_config(pipeline.scheduler.config)

        for _ in pipeline(prompt='A gorgeous smiling slim young japanese girl, befautiful face, hands with five fingers, light background, best quality, extremely detailed, HD, ultra-realistic, 8K, HQ, masterpiece, trending on artstation, art, smooth', frames=frames, control_frames=openpose, generator=torch.manual_seed(0), strength=strength[index], negative_prompt='monochrome, dark background, longbody, lowres, bad anatomy, bad hands, fused fingers, missing fingers, too many fingers, extra digit, fewer difits, cropped, worst quality, low quality, deformed body, bloated, ugly, unrealistic, extra hands and arms').frames: writer.append_data(numpy.asarray(_))

if __name__ == '__main__': torch_xla.launch(process)

You see the origin version require two long lists of PIL.Image before hand. one for frames and one for control_frames. Now the control_frames can be calculated on the fly which only need one list of PIL.Image before hand. so the cpu memory usage can be half compare to the original version