Skip to content

Latest commit

 

History

History
189 lines (99 loc) · 7.49 KB

user_faq.md

File metadata and controls

189 lines (99 loc) · 7.49 KB

Why does it require a graphic accelerator to work?

The program is based on machine learning models. For example, face detection in the frame, facial point detection, face replacement. These models require intensive computations, which graphical accelerator can handle tens of times faster than on an ordinary processor.

How to increase the quality of the picture?

Play with different program settings, try different lighting in the room, a different camera angle.

Why doesn't the replaced face look like a celebrity?

It depends on how much your face fits the shape of the celebrity's face.

Why is the swapped face blurrier than the rest of the image?

It depends on how big the face in the frame, as well as the resolution of the model. The public models have a resolution of 224x224 to work on most PC configurations. For high face resolution, you need to train your own model. Public models are suitable for game streamers that have their face in a small window in the corner of the screen.

Can I play at the same time as DeepFaceLive is running?

It depends on how powerful your computer and how demanding the game to resources. A multi-core processor and a separate graphics card for face replacement are recommended.

How to increase the performance of face replacement?

Play with different program settings. Any module put on the CPU will consume a lot of CPU time, which is not enough to run a game, for example. If the motherboard allows, you can install additional video cards and distribute the load on them. I also suggest watching the CPU load in Task Manager while experimenting with the settings.

Will DeepFake detectors be able to detect a fake in my streams?

depends on final quality of your picture. Flickering face, abruptly clipping face mask, irregular color will increase chance to detect the fake.

deepware.ai detects no fakes in the example video of the fake Margot Robbie.

Do I need to train the model from scratch?

No you don't. There are public face models that can swap any face without training.

I want to have more control when changing faces in a video. Will the new functionality be implemented?

No. DeepFaceLive is designed for face swapping in streams. The ability to change faces in the videos - only for test purposes.

I want to swap my face to a particular celebrity. What I need to do?

If you are novice, learn all about DeepFaceLab https://mrdeepfakes.com/forums/thread-guide-deepfacelab-2-0-guide

Gather 5000+ samples of your face with various conditions using webcam which will be used for Live. The conditions are as follows: different lighting, different facial expressions, head direction, eyes direction, being far or closer to the camera, etc. Sort faceset by best to 2000.

Here public storage https://disk.yandex.ru/d/7i5XTKIKVg5UUg with facesets and models.

Using pretrained "RTT model 224 V2.zip" from public storage (see above)

Make a backup before every stage !

  1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned

  2. place your celeb to workspace/data_src/aligned

  3. do not change settings. Train +500.000

  4. replace dst faceset with your faceset in workspace/data_dst/aligned

  5. continue train +500.000, (optional) deleting inter_AB.npy every 100.000 (save, delete, continue run)

  6. random_warp:OFF, GAN 0.1 power, patch size 28, gan_dims:32. Train +700.000

Using SAEHD model from scratch.

res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:N, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.

Make a backup before every stage !

  1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned

  2. place your celeb to workspace/data_src/aligned

  3. train +1.000.000 deleting inter_AB.npy every 100.000 (save, delete, continue run)

  4. place your faceset to workspace/data_dst/aligned

  5. do not delete anything, continue train +500.000

  6. random_warp:OFF, GAN 0.1 power, patch size 28, gan_dims:32. Train +700.000

  7. export the model in .dfm format for use in DeepFaceLive. You can also try ordering a deepfake model from someone in Discord or forum.

I want to train ready-to-use face model to swap any face to celebrity, same as public face model. What I need to do?

If you are familiar with DeepFaceLab, then this tutorial will help you:

Src faceset is celebrity. Must be diverse enough in yaw, light and shadow conditions. Do not mix different age. The best result is obtained when the face is filmed from a short period of time and does not change the makeup and structure. Src faceset should be xseg'ed and applied. You can apply Generic XSeg to src faceset.

Using pretrained "RTT model 224 V2.zip" from public storage (see above)

Make a backup before every stage !

  1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned

  2. place your celeb to workspace/data_src/aligned

  3. place model folder to workspace/model

  4. do not change settings, train +500.000 iters, + deleting inter_AB.npy every 100.000 (save, delete, continue run)

  5. random_warp:OFF, GAN 0.1 power, patch size 28, gan_dims:32. Train +700.000

Using SAEHD model from scratch

res:224, WF, archi:liae-udt, ae_dims:512, e_dims:64, d_dims:64, d_mask_dims:32, eyes_mouth_prio:N, blur_out_mask:Y, uniform_yaw:Y, lr_dropout:Y, batch:8. Others by default.

Make a backup before every stage !

  1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned

  2. place your celeb to workspace/data_src/aligned

  3. train +2.000.000 iters, deleting inter_AB.npy every 100.000-500.000 iters (save, delete, continue run)

  4. random_warp still ON, train +500.000

  5. random_warp:OFF, GAN 0.1 power, patch size 28, gan_dims:32. Train +700.000

reusing trained SAEHD RTM model

Models that are trained before random_warp:OFF, can be reused. In this case you have to delete INTER_AB.NPY from the model folder and continue training from stage where random_warp:ON. Increase stage up to 2.000.000 and more iters. You can delete inter_AB.npy every 500.000 iters to increase src-likeness. Trained model before random_warp:OFF also can be reused for new celeb face.

Is there any way to train a model faster?

You can train a model in 1 day on RTX 3090, sacrificing quality.

1-day-3090 training. Using pretrained "RTT model 224 V2.zip" from public storage (see above)

  1. place RTM WF Faceset V2 from public storage (see above) to workspace/data_dst/aligned

  2. place your celeb to workspace/data_src/aligned

  3. place model folder to workspace/model

  4. do not change settings, train +25.000 iters

  5. delete inter_AB.npy (save, delete, continue run)

  6. train +30.000 iters

  7. random_warp OFF, GAN 0.1 power, patch size 28, gan_dims:32.

  8. Turn off after 24 hours

I want to change some code and test the result on my local machine. What I need to do?

There is ready-to-use VSCode editor inside DeepFaceLive folder located in

_internal\vscode.bat

All code changes will only affect the current folder.

Also you can build a new and clean DeepFaceLive folder with the code from current folder using

_internal\build DeepFaceLive.bat