Add link to mDiffAE v2 as recommended version
Browse files
README.md
CHANGED
|
@@ -11,6 +11,13 @@ library_name: mdiffae
|
|
| 11 |
|
| 12 |
# mdiffae_v1
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
**mDiffAE** — **M**asked **Diff**usion **A**uto**E**ncoder.
|
| 15 |
A fast, single-GPU-trainable diffusion autoencoder with a **64-channel**
|
| 16 |
spatial bottleneck. Uses decoder token masking as an implicit regularizer
|
|
|
|
| 11 |
|
| 12 |
# mdiffae_v1
|
| 13 |
|
| 14 |
+
> **[mDiffAE v2](https://huggingface.co/data-archetype/mdiffae-v2) is now available and is the recommended version.** It offers substantially better reconstruction (+1.7 dB mean PSNR) with the same or better downstream convergence.
|
| 15 |
+
>
|
| 16 |
+
> | Version | Mean PSNR (2k images) | Bottleneck | Decoder |
|
| 17 |
+
> |---|---|---|---|
|
| 18 |
+
> | [**mDiffAE v2**](https://huggingface.co/data-archetype/mdiffae-v2) (recommended) | **35.81 dB** | 96ch (8x) | 8 blocks (skip-concat) |
|
| 19 |
+
> | mDiffAE v1 (this repo) | 34.15 dB | 64ch (12x) | 4 blocks (flat) |
|
| 20 |
+
|
| 21 |
**mDiffAE** — **M**asked **Diff**usion **A**uto**E**ncoder.
|
| 22 |
A fast, single-GPU-trainable diffusion autoencoder with a **64-channel**
|
| 23 |
spatial bottleneck. Uses decoder token masking as an implicit regularizer
|