Feature Suggestion - include latent_directions

[johndpope]

So I've been trying out different stylegan repos - and I'm imaging a fusion of two. Basically the guts of this repo https://github.com/GreenLimeSia/GenEdi/issues/3

While it is in pytorch - I think it would be great to surface these in side bar as in the 512 vectors currently in another tab.

Also checkout artbreeder https://www.artbreeder.com/ it has some simple controls to switch up / infuse latent vectors.

Side note - the nvidia labs are switching over to pytorch - https://github.com/NVlabs/stylegan2-ada/issues/32#issuecomment-748225014

[k-l-lambda]

Nice work.

@johndpope, I guess you have tried this GenEdi, how accurate are these feature vectors you think?

[johndpope]

I think these are just out of the box. https://github.com/GreenLimeSia/GenEdi/tree/master/latent_directions

https://github.com/anvoynov/GANLatentDiscovery

https://github.com/GreenLimeSia/GenEdi/blob/master/latent_directions/eyes_open.npy

https://github.com/search?q=latent_directions&type=code

@Gvanderl - maybe there's interest to collaborate with this web layout for your facemaker code? https://github.com/Gvanderl/FaceMaker

UPDATE also noteworthy @blubs https://github.com/blubs/stylegan2_playground

[k-l-lambda]

Very nice. I'm curious how these latent directions been got. StyleGAN is not trained with feature labels, latent space is very likely highly twisted. So did these latent directions come from manual labeling or unsupervised learning?

[johndpope]

I think this guy @a312863063 created them https://github.com/a312863063/generators-with-stylegan2

english https://github.com/a312863063/generators-with-stylegan2/blob/master/README_EN.md

[k-l-lambda]

Ha, I know some about this guy. He has a lot of computation resource, but the researching is not the first-class yet. Yes, I envy him.

[johndpope]

Amazon / stylegan2 take Gan-Control https://arxiv.org/pdf/2101.02477v1.pdf

https://alonshoshan10.github.io/gan_control/

[k-l-lambda]

Amazon / stylegan2 take Gan-Control https://arxiv.org/pdf/2101.02477v1.pdf

https://alonshoshan10.github.io/gan_control/

Interesting.

[johndpope]

related - https://github.com/johndpope/ALAE/blob/master/interactive_demo.py

[gael-vanderlee]

I think these are just out of the box. https://github.com/GreenLimeSia/GenEdi/tree/master/latent_directions

https://github.com/anvoynov/GANLatentDiscovery

https://github.com/GreenLimeSia/GenEdi/blob/master/latent_directions/eyes_open.npy

https://github.com/search?q=latent_directions&type=code

@Gvanderl - maybe there's interest to collaborate with this web layout for your facemaker code? https://github.com/Gvanderl/FaceMaker

UPDATE also noteworthy @blubs https://github.com/blubs/stylegan2_playground

I would love to collaborate, I'm not certain what exactly should be implemented though.

[johndpope]

Hi @Gvanderl nice repo! https://github.com/Gvanderl/StyleSegments

There seems to be 3 parts - (though maybe more) Python API

• expose the directory contents as file names via api / could be done via the initial spec call
• DRAFTED

{ "image_shape": [ null, 3, 1024, 1024 ], "latent_directions": [ "latent_directions/emotion_angry.npy" 1: "latent_directions/angle_vertical.npy" 2: "latent_directions/emotion_fear.npy" 3: "latent_directions/lip_ratio.npy" 4: "latent_directions/pitch.npy" 5: "latent_directions/exposure.npy" 6: "latent_directions/roll.npy" 7: "latent_directions/eyes_open.npy" 8: "latent_directions/beauty.npy" 9: "latent_directions/nose_ratio.npy" 10: "latent_directions/glasses.npy" 11: "latent_directions/eye_eyebrow_distance.npy" 12: "latent_directions/face_shape.npy" 13: "latent_directions/mouth_open.npy" 14: "latent_directions/nose_tip.npy" 15: "latent_directions/eye_distance.npy" 16: "latent_directions/race_yellow.npy" 17: "latent_directions/mouth_ratio.npy" 18: "latent_directions/smile.npy" 19: "latent_directions/emotion_surprise.npy" 20: "latent_directions/race_black.npy" 21: "latent_directions/angle_horizontal.npy" 22: "latent_directions/gender.npy" 23: "latent_directions/emotion_happy.npy" 24: "latent_directions/race_white.npy" 25: "latent_directions/width.npy" 26: "latent_directions/emotion_disgust.npy" 27: "latent_directions/camera_rotation.npy" 28: "latent_directions/age.npy" 29: "latent_directions/height.npy" 30: "latent_directions/yaw.npy" 31: "latent_directions/nose_mouth_distance.npy" 32: "latent_directions/emotion_sad.npy" 33: "latent_directions/eye_ratio.npy" 34: "latent_directions/emotion_easy.npy" ], "latents_dimensions": 512, "model": "cat", "synthesis_input_shape": [ null, 18, 512 ] }

ML
- when app changes the values - make them update backend
VUE.js
-surface the latent_directions with sliders

1st step import the latent_directions directory - DONE

I pushed a tensorflow 2 branch. Work in progress.
https://github.com/johndpope/stylegan-web

It seems each latent directions is a vector of  'shape': (18, 512)

That may makes things more complicated.
I'm thinking we could have a drop down - then when they select - show the 512 latent vectors associated with say age. 

This looks promising - need to prototype the code to hotwire the http_server (without UI)

```python
def change_face(image="maface_01", direction="gender", coeffs=None):

    if coeffs is None:
        coeffs = [-2, 0, 2]

    directions = {
        "smile": 'ffhq_dataset/latent_directions/smile.npy',
        "gender": 'ffhq_dataset/latent_directions/gender.npy',
        "age": 'ffhq_dataset/latent_directions/age.npy'
    }
    direction = np.load(directions[direction])
    face_latent = np.load(config.latents_dir / (image + ".npy"))
    move_and_show(face_latent, direction, coeffs)

If we can hardcode a change of smile - then this would progress things to extend ui.

UPDATE @Puzer Dmitry Nikitko originally found directions - according to this tweet UPDATE - they're different / I've added them all in.

https://hostb.org/NCM - official download here https://twitter.com/robertluxemburg/status/1207087801344372736 https://github.com/Puzer/stylegan-encoder/blob/master/Learn_direction_in_latent_space.ipynb

UPDATE.
I added a killServer.sh to take out the http_server and a restart.sh script that also runs the build to recompile vuejs code.

Got the drop down showing latent directions in project.

It seems final piece would be to wire up the events of changing the drop down - to loading the corresponding pkl and then making this become the generator??? I've wired up the event onChange - so it calls the backend and selects to load the respective pkl / needs work here. The generator would either need to take a turn to include a selected blended model - or ....IDK.

UPDATE seems like we can just concatenate the pre-canned latent directions with some simple scaling as done by @jasonlbx13

https://github.com/jasonlbx13/FaceHack/blob/master/face_gan/flask_app.py

  if len(request.form) != 0:
                    smile = float(request.form['smile'])
                    age = float(request.form['age'])
                    gender = float(request.form['gender'])
                    beauty = float(request.form['beauty'])
                    angleh = float(request.form['angleh'])
                    anglep = float(request.form['anglep'])
                    raceblack = float(request.form['raceblack'])
                    raceyellow = float(request.form['raceyellow'])
                    racewhite = float(request.form['racewhite'])
                    feature_book = [smile, age, gender, beauty, angleh, anglep, raceblack, raceyellow, racewhite]
                else:
                    feature_book = [0, 0, 0, 0, 0, 0, 0, 0, 0]

def move_latent(self, npy_dir, Gs_network, Gs_syn_kwargs, *args):
        latent_vector = np.load(npy_dir)[np.newaxis, :]
        smile, age, gender, beauty, angleh, anglep, raceblack, raceyellow, racewhite = args
        new_latent_vector = latent_vector.copy()
        new_latent_vector[0][:8] = (latent_vector[0] + smile * self.smile_drt + age * self.age_drt + gender * self.gender_drt
                                    + beauty * self.beauty_drt + angleh * self.angleh_drt + anglep * self.anglep_drt
                                    + raceblack * self.raceblack_drt + raceyellow * self.raceyellow_drt + racewhite * self.racewhite_drt)[:8]
        with self.graph.as_default():
            with self.session.as_default():
                images = Gs_network.components.synthesis.run(new_latent_vector, **Gs_syn_kwargs)
        PIL.Image.fromarray(images[0], 'RGB').save(
            dnnlib.make_run_dir_path('./static/img/edit_face.jpg'))

I guess once code is dropped in - the coefficients will be important to get this to work correctly. https://github.com/dream80/StyleganEncoder/blob/d9c8fe41a5b3ccdebd6af9d234a45c97416af3af/show1.py

#move_and_show(generator, target, "age", [-6, -4, -3, -2, 0, 2, 3, 4, 6])
#move_and_show(generator, target, "angle_horizontal", [-6, -4, -3, -2, 0, 2, 3, 4, 6])
#move_and_show(generator, target, gender, [-6, -4, -3, -2, 0, 2, 3, 4, 6])
#move_and_show(generator, target, "eyes_open", [-3, -2, -1, -0.5, 0, 0.5, 1, 2, 3])
#move_and_show(generator,target, glasses, [-6, -4, -3, -2, 0, 2, 3, 4, 6])
#move_and_show(generator,target, smile, [-3, -2, -1, -0.5, 0, 0.5, 1, 2, 3])

Also noteable mention @swapp1990 - seems to have achieved something similar already. https://github.com/swapp1990/my-stylegan2-face-editing-app/blob/dev/client/src/views/FaceEditing.vue