Zielon / INSTA

INSTA - Instant Volumetric Head Avatars [CVPR2023]
https://zielon.github.io/insta/
Other
438 stars 37 forks source link

Can I use the existing code to do expression transfer #6

Closed lishaoxu1994 closed 1 year ago

lishaoxu1994 commented 1 year ago

Thanks for your solution for headless trainging, it works well. In the Supplemental Document->B. Applications, the expression transfer results are attractive. Is it possible to do expression transfer with the existing code?

Zielon commented 1 year ago

Yes, the current code supports it, however, I did not upload the relative expression generation script. I maybe release this part as well. You just need to have the corresponding transforms.json file and meshes.

lishaoxu1994 commented 1 year ago

Yes, the current code supports it, however, I did not upload the relative expression generation script. I maybe release this part as well. You just need to have the corresponding transforms.json file and meshes.

thanks for your answer. I also wanna know how to render images with existing checkpoints.

Zielon commented 1 year ago

You can use "render_from_snapshot": true # For --no-gui option to directly render sequences from the config.

lishaoxu1994 commented 1 year ago

You can use "render_from_snapshot": true # For --no-gui option to directly render sequences from the config.

I use "render_from_snapshot":true, and set the --scene, --no-gui, --snapshot, It seems nothing happens.

And after I train a scene, the max_steps is fixed. I change the max_steps and use --snapshot, the traing shows the initial max_steps.

Zielon commented 1 year ago

You are right there was a bug. Now it is fixed, you can also specify render_novel_trajectory option in this mode.

lishaoxu1994 commented 1 year ago

You are right there was a bug. Now it is fixed, you can also specify render_novel_trajectory option in this mode.

It doesn't seem to work.

Zielon commented 1 year ago

I am using:

--config insta.json --scene data/obama/transforms_test.json --snapshot data/obama/snapshot.msgpack --no-gui

{
    "parent": "main.json",
    "max_steps": 33000,
    "max_cached_bvh": 4000,
    "max_images_gpu": 1700,
    "use_dataset_cache": true,
    "render_novel_trajectory": false,
    "render_from_snapshot": true
}

The output should be in ./data/obama/experiments/insta/debug/overlay if you need more trajectories use "render_novel_trajectory": true

lishaoxu1994 commented 1 year ago

I am using:

--config insta.json --scene data/obama/transforms_test.json --snapshot data/obama/snapshot.msgpack --no-gui

{
  "parent": "main.json",
  "max_steps": 33000,
  "max_cached_bvh": 4000,
  "max_images_gpu": 1700,
  "use_dataset_cache": true,
  "render_novel_trajectory": false,
  "render_from_snapshot": true
}

The output should be in ./data/obama/experiments/insta/debug/overlay if you need more trajectories use "render_novel_trajectory": true

I use the same config, the code only outputs after training. When I try to render reults again, nothing appears.

Zielon commented 1 year ago

Hmm, have you pulled the newest changes and rebuilt the project?

alchemician commented 1 year ago

@lsx0101 it is just replacing the previous files in the debug/overlay folder

lishaoxu1994 commented 1 year ago

@lsx0101 it is just replacing the previous files in the debug/overlay folder

it works, but the rendering quality is not as good as training expressions

oijoijcoiejoijce commented 1 year ago

I'm gettin these results; imperfect rendering and transfer, is anyone getting better output?

https://github.com/Zielon/INSTA/assets/58717195/ea96b423-1bf1-420b-947b-30fe6fd1026a