The new model is fully created and trained (for animation). I'm still need to finish training it for RL, most likely will be complete sometime next week.
Overall, it must have take around 400~500h of training, but mostly of it was failed tests. The final model must had take around ~40h of training.
I am my own worst critic, so its hard to say if the results are good or bad, will let you fellows be the judge once it's released, but for a first time creating a model, I th...
2021-09-30 13:22:42 +0000 UTC
View Post
I trained a partial model for RL, and the results are good, but I feel that they still can be improved, so I'm still tampering with it.
Real Life pretty much also require "Step 2" of the model to generate good results, so while I'm making the second part of the model, I can keep testing stuff with the RL model.
2021-09-27 19:12:40 +0000 UTC
View Post
2021-09-24 00:23:14 +0000 UTC
View Post
2021-09-24 00:22:34 +0000 UTC
View Post
2021-09-24 00:21:50 +0000 UTC
View Post
2021-09-24 00:21:16 +0000 UTC
View Post
Ok, some stuff first:
- I found a bug in the model, so I'm training again. On top of that I'm trying some extra stuff that I been studying, so tomorrow I may have a better model? Hard to say with AI.
- This examples are using only Step 1 of the model. There is some advantages and dis some good and bad on that.
Good:
- Using only Step 1 you can interpolate 2X 3X 4X 5X 6X 7X 8X [..] 99X with almost no extra time. The extra time is pretty m...
2021-09-24 00:20:34 +0000 UTC
View Post
The first part of my new model is complete and giving some decent results.
It's trained on animation and anime, but while I write this, i'm training one for real life footage.
While the results are pretty decent, I'm kinda of crazy and can't never be pleased with it give, so most likely I will try to keep improving it lol
Creating and training a model was pretty fun and made me learn a lot of new stuff, stuff that now I can use to improve all my projects.
I...
2021-09-23 12:16:04 +0000 UTC
View Post
Hi there, here is a small list of stuff I'm working right now. In no particular order.
- Fixing and adding a bunch of stuff that was requested on Rife-App
- Rife-App Training is looking good, I'm focusing on animation, but there is no reason it won't work with real life stuff, just need to train a separated model, hope its work once it's finished. As long the results are good, I plan to release a public Git in the future with all the stuff to create motion flows for animation...
2021-09-17 14:51:00 +0000 UTC
View Post
Me: Training a model for 20h.
What happen if I change this value?
New model a little better than the last.
Hooo. Start from scratch.
2021-09-15 19:39:23 +0000 UTC
View Post
I must had put 300h/400h of training testing a lot of my theories on interpolation, results was alright, but worse than Rife. I pretty much had give up on training, for now at least.
I just left the last few tests run til the end. In my last test I managed to get some pretty decent results. lol
It's a self made model, using as inspiration Rife. I need to revisit Dain to learn a little more of its algorithm.
So now I'm giving this result some more training and some other tests ...
2021-09-14 20:40:35 +0000 UTC
View Post
This pool don't mean I will stop working on stuff like Rife-App and Clip-App, but there is some apps that been some time since I made an update, like FlowBlur.
So I want to see if there is any desire to keep updating one of those less popular projects.
(You can select more than one option)
2021-09-08 13:04:43 +0000 UTC
View Post
A few days ago someone suggested to me to create a Real-ESRGAN interface. It was a great idea, I talked with the Real-ESRGAN creator and he said it was alright. So here it is.
Ok, first thing, this is a rushed release, but I would need around 3 days to fix the small bugs, so here is a rushed release so you guys and gals don't need to wait a few more days.
The app is working with image and videos, but a few pointers:
- There is 3 models: 2X /...
2021-09-07 01:00:14 +0000 UTC
View Post
It's annoying on how much fun is trying to improve the interpolation stuff. To bad is so time consuming.
Anyway, I think the the new app will be ready today, there was a update today on the main code that I want to implement as well, so it was a great timed update. Just hope it's easy to integrate to my working code.
There is also a Clip-App update on the oven, but still need some work before it can be updated.
After this new app is released, I will make a pool to see what e...
2021-09-06 14:18:33 +0000 UTC
View Post
Yes, very useful app, very mysterious.
2021-08-31 20:03:17 +0000 UTC
View Post
Here a test with the last model. You can slow down the player to check the details.
2021-08-26 01:17:04 +0000 UTC
View Post
Here it is, the final animation model (for now).
There still a lot to be improved, if the model fail to interpolate, it can generate pretty ugly stuff, plus it could probably work a little better in some situations, but overall I think is giving good results. If you are doing HD stuff, putting the scale to 0.5 may improve a little if the results are bad with scale 1
2021-08-26 01:14:51 +0000 UTC
View Post
Right now I'm training a new Anime model for Rife, I think this one will be the best one yet.
Also working on the new version of Clip-App with some extra options and a new application. It should be ready soon.
2021-08-23 11:50:31 +0000 UTC
View Post
Will see if I can implement on the GUI soon. Check it out.
https://twitter.com/ViktorAlm/status/1428095004963229700
https://twitter.com/dribnet
And here a working colab for now.
ht...
2021-08-19 13:27:55 +0000 UTC
View Post
Ok, let me start saying that the Anime model probably can get better, but it's already usable in the current state.
Also, there is the need to redo the GUI a little, I messed up and now the GUI is a little bigger than the screen. You need to scale down a little the window to see the Render button. Will fix on the next update.
- There is now a option to select the PNG encode, at 0 it will not encode the PNG files, it will write them faster but it will also take more ...
2021-08-19 00:49:52 +0000 UTC
View Post
Overall, the results are very good. I'm enjoying the results I'm getting from my tests. It still generate artifacts if the movements are very crazy.
Could it be improved further? Probably, but I need to upgrade my rig or rend a GPU service, I can't keep making my computer stuck only with those stuff, gonna need to do some thinking before training it more.
I still fixing some stuff on the app, but I hope tomorrow I can upload the new Rife-App with this new model.
2021-08-18 01:28:27 +0000 UTC
View Post
It work well with Real Footage, but I still did not test if the results are better or worse than the older models.
2021-08-18 01:25:39 +0000 UTC
View Post
It fix some artifacts from the models, there still a few artifacts close to the end.
2021-08-18 01:24:38 +0000 UTC
View Post
By naked eye it may not change a lot from the other versions, but on my smaller tests, this model seen to work a lot better for animations.
2021-08-18 01:23:26 +0000 UTC
View Post
I still want to make it be able to generate a somewhat decent pixel-art. I think it would be nice for text games/ Table RPG games / overall indie games. Perhaps someday it could work directly inside Unity or other engine. Someday...
2021-08-17 02:27:40 +0000 UTC
View Post
███████████████████████████████████████████████
Hi all, for the last 6 or 7 days my computer been really busy with Rife-App.
I been trying to train the AI to work better with animation/anime. I don't like to promise that the model will work better, because the thing with training a model is that you may train it for weeks and in the end it will be garbage. But i'm feeling the one i'm trai...
2021-08-16 22:04:22 +0000 UTC
View Post
First thing. I forgot to update the version inside the app, so it still say 1.0 lol
There was some memory optimization and now the folders are generated with the correct seed, not -1
1. Kornia Alternative: Is Kornia with some diff settings, it can generate different results.
2. Layers: This is the number of layers to be used by the AI. 64 is the default number and a good value, but ...
2021-08-11 23:28:50 +0000 UTC
View Post
Using a image as a starting point, Clip try to morph that image on the text you require.
Origin:
The Stars in the cosmos
The build should be ready later today.
2021-08-11 17:27:24 +0000 UTC
View Post
I would like to release the new version of Rife-APP first, but I'm currently testing some stuff with it what I still was unable to complete, so Clip-App will come first.
I'm adding some new options, like target and source images, and a new Image model. Also improved the memory usage a little. More information once it's out.
2021-08-10 22:47:10 +0000 UTC
View Post
Here we go, the app from the previous post.
As seen in the previous examples, this AI try to mask people from photos or videos.
It is based on this git:
https://github.com/ZHKKKe/MODNet
And it have 3 export modes:
RGBA: It generate a PNG with a alpha channel
RGB + Alpha PNG: A regular RGB Png and another PNG with the alpha as RGB
GreenScreen: A Mp4 with the mask as a green color....
2021-08-07 01:42:33 +0000 UTC
View Post
