README: collapse Performance section, add badges for all model-zoo repos
Browse files
README.md
CHANGED
|
@@ -15,10 +15,18 @@ tags:
|
|
| 15 |
|
| 16 |
<p align="center">
|
| 17 |
<a href="https://github.com/microsoft/Mage"><img src="https://img.shields.io/badge/Code-GitHub-181717?logo=github" alt="GitHub"></a>
|
| 18 |
-
<a href="https://huggingface.co/microsoft/Mage-Flow-Base"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-ffcc00" alt="Hugging Face"></a>
|
| 19 |
<a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-green" alt="License: MIT"></a>
|
| 20 |
</p>
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
<div align="center">
|
| 23 |
<img src="assets/mage-flow-cover.png" width="100%" alt="gallery">
|
| 24 |
</div>
|
|
@@ -140,6 +148,9 @@ Each checkpoint is a self-contained diffusers-style repo (`transformer/` + share
|
|
| 140 |
|
| 141 |
## 📊 Performance
|
| 142 |
|
|
|
|
|
|
|
|
|
|
| 143 |
**Text-to-image** — full benchmark suite: GenEval, DPG-Bench, TIIF-Bench (short/long splits), CVTG-2K, OneIG (EN/CN), LongText (EN/CN). Higher is better; GenEval / CVTG-2K / OneIG / LongText on a 0–1 scale, DPG / TIIF on 0–100. The **Type** column marks closed- vs open-source; **bold** / <ins>underline</ins> = best / second-best among **open-source** models (closed-source shown for reference, not ranked); `–` = not reported; ★ = ours.
|
| 144 |
|
| 145 |
| Model | Type | #Params | Steps | GenEval | DPG | TIIF-Short | TIIF-Long | CVTG-2K | OneIG-EN | OneIG-CN | LongText-EN | LongText-CN |
|
|
@@ -190,6 +201,8 @@ Each checkpoint is a self-contained diffusers-style repo (`transformer/` + share
|
|
| 190 |
| **Mage-Flow-Edit** ★ | Open | 4B | 30 | 4.34 | 8.127 | 8.123 | 14.14 | 16.26 |
|
| 191 |
| **Mage-Flow-Edit-Turbo** ★ | Open | 4B | 4 | 4.38 | <ins>8.271</ins> | **8.264** | 12.77 | 15.41 |
|
| 192 |
|
|
|
|
|
|
|
| 193 |
## 🏗️ Architecture
|
| 194 |
|
| 195 |
**Mage-VAE** — a latent tokenizer built as a *symmetric* one-step diffusion codec: the decoder is a fully-convolutional one-step pixel-diffusion model (no global-attention blocks), and the encoder is its architectural dual (a one-step latent generator conditioned on pixels). A standard Gaussian-prior KL is replaced with an **anchor-latent KL** that regularizes the posterior toward FLUX.2-VAE latents, giving a generation-ready `128`-channel, `16×`-downsampled latent space.
|
|
|
|
| 15 |
|
| 16 |
<p align="center">
|
| 17 |
<a href="https://github.com/microsoft/Mage"><img src="https://img.shields.io/badge/Code-GitHub-181717?logo=github" alt="GitHub"></a>
|
|
|
|
| 18 |
<a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-green" alt="License: MIT"></a>
|
| 19 |
</p>
|
| 20 |
|
| 21 |
+
<p align="center">
|
| 22 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow-Base"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow--Base-ffcc00" alt="Mage-Flow-Base"></a>
|
| 23 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow-ffcc00" alt="Mage-Flow"></a>
|
| 24 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow-Turbo"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow--Turbo-ffcc00" alt="Mage-Flow-Turbo"></a>
|
| 25 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow-Edit-Base"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow--Edit--Base-ff9900" alt="Mage-Flow-Edit-Base"></a>
|
| 26 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow-Edit"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow--Edit-ff9900" alt="Mage-Flow-Edit"></a>
|
| 27 |
+
<a href="https://huggingface.co/microsoft/Mage-Flow-Edit-Turbo"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Mage--Flow--Edit--Turbo-ff9900" alt="Mage-Flow-Edit-Turbo"></a>
|
| 28 |
+
</p>
|
| 29 |
+
|
| 30 |
<div align="center">
|
| 31 |
<img src="assets/mage-flow-cover.png" width="100%" alt="gallery">
|
| 32 |
</div>
|
|
|
|
| 148 |
|
| 149 |
## 📊 Performance
|
| 150 |
|
| 151 |
+
<details>
|
| 152 |
+
<summary><b>Full benchmark tables (text-to-image & image editing) — click to expand</b></summary>
|
| 153 |
+
|
| 154 |
**Text-to-image** — full benchmark suite: GenEval, DPG-Bench, TIIF-Bench (short/long splits), CVTG-2K, OneIG (EN/CN), LongText (EN/CN). Higher is better; GenEval / CVTG-2K / OneIG / LongText on a 0–1 scale, DPG / TIIF on 0–100. The **Type** column marks closed- vs open-source; **bold** / <ins>underline</ins> = best / second-best among **open-source** models (closed-source shown for reference, not ranked); `–` = not reported; ★ = ours.
|
| 155 |
|
| 156 |
| Model | Type | #Params | Steps | GenEval | DPG | TIIF-Short | TIIF-Long | CVTG-2K | OneIG-EN | OneIG-CN | LongText-EN | LongText-CN |
|
|
|
|
| 201 |
| **Mage-Flow-Edit** ★ | Open | 4B | 30 | 4.34 | 8.127 | 8.123 | 14.14 | 16.26 |
|
| 202 |
| **Mage-Flow-Edit-Turbo** ★ | Open | 4B | 4 | 4.38 | <ins>8.271</ins> | **8.264** | 12.77 | 15.41 |
|
| 203 |
|
| 204 |
+
</details>
|
| 205 |
+
|
| 206 |
## 🏗️ Architecture
|
| 207 |
|
| 208 |
**Mage-VAE** — a latent tokenizer built as a *symmetric* one-step diffusion codec: the decoder is a fully-convolutional one-step pixel-diffusion model (no global-attention blocks), and the encoder is its architectural dual (a one-step latent generator conditioned on pixels). A standard Gaussian-prior KL is replaced with an **anchor-latent KL** that regularizes the posterior toward FLUX.2-VAE latents, giving a generation-ready `128`-channel, `16×`-downsampled latent space.
|