-
hi @lllyasviel amazing work, i tried replicating the training strategy in the paper, and was able to get a model working, but the results i see are cloudy and have artifacts sometimes, lack consisten…
-
In looking at manifolds in my own experiments, I have noticed a consistent "dead zone" near the origin of the latent space. Here is an example generated with `faces/train_uncond_dcgan.py` and z=100:
…
-
The Iterative Latent Upscale utilizes the function `latent_upscale_on_pixel_space`, which does the following:
1. decode the input latent into pixel space
2. scale the pixels
3. encode the scaled pi…
-
Prompt interpolation currently uses embeddings to calculate values in-between control points. I'm not sure whether this is the right way to interpolate prompts, or if we should instead work in latent …
-
When running a uniform write workload, on a uniformly sized keyspace, ranges will split in half at approximately the same time.
Normally, when a few ranges split this isn't expected to have any im…
kvoli updated
1 month ago
-
So, I'm not bright. Can anyone who is brighter than me explain what "number of repeats in CLIP latent space" is for / does? Thanks!
-
## 一言でいうと
VAEを学習するとき、潜在表現zが無視され事前分布と事後分布を一致させる方向に学習が行われてしまうことがある(KL collapse)。これを解消するために、正規分布の代わりにVon Mises–Fisher分布(球面上の分布)を導入。KL距離からパラメーター要素(μ,σ)を除去することで、学習による一致を防いでいる
![image](https://user-im…
-
## 一言でいうと
GANはCNN+Adversarialな学習という点に特徴があるが、後者の形態は学習を難しくしている。そこでdiscriminator(以下D)をとってしまい、ノイズとサンプルから元画像を復元する。そしてDの持つ判定力を模倣するため、ラプラシアン階層ごとの差分を利用し学習する。
### 論文リンク
https://arxiv.org/abs/1707.0577…
-
Hello, I'm wondering why the dimensionality of the latent W is (1,16,512) instead of (1,18,512) which is proposed by the original StyleGAN2. I'm asking because I'm trying to edit real images which are…
-
Is the visualization in pixel space or latent space?