multimodalart HF Staff commited on
Commit
ef8b23f
·
verified ·
1 Parent(s): e04067c

Apply Qwen-Image-Edit Lightning 8-step LoRA

Browse files
Files changed (1) hide show
  1. README.md +14 -6
README.md CHANGED
@@ -1,6 +1,6 @@
1
  ---
2
- title: MetaView Novel View Synthesis
3
- emoji: 🎥
4
  colorFrom: pink
5
  colorTo: purple
6
  sdk: gradio
@@ -8,25 +8,33 @@ sdk_version: 5.49.1
8
  app_file: app.py
9
  pinned: false
10
  license: apache-2.0
11
- short_description: Monocular novel view synthesis from a single image
12
  python_version: "3.10"
13
  startup_duration_timeout: 1h
14
  ---
15
 
16
- # MetaView — Monocular Novel View Synthesis
17
 
18
  Interactive demo for [**MetaView**](https://huggingface.co/Kwai-Kolors/MetaView)
19
  (ECCV 2026): *Monocular Novel View Synthesis with Scale-Aware Implicit Geometry
20
- Priors*.
21
 
22
  Upload a single image, choose a target camera **yaw** (left/right) and **pitch**
23
- (up/down), and MetaView renders the scene from the new viewpoint.
 
24
 
25
  The pipeline combines:
26
  - **Qwen-Image-Edit** MM-DiT backbone (novel-view generation),
 
 
27
  - **Depth-Anything-3** GIANT (implicit 3D feature priors) + NESTED (dense depth),
28
  - **Modified RoPE (PRoPE)** for metric scale anchoring of the camera pose.
29
 
 
 
 
 
 
30
  Links: [Model](https://huggingface.co/Kwai-Kolors/MetaView) ·
31
  [Code](https://github.com/KlingAIResearch/MetaView) ·
32
  [Paper](https://arxiv.org/abs/2607.12000)
 
1
  ---
2
+ title: MetaView Lightning Novel View Synthesis
3
+ emoji:
4
  colorFrom: pink
5
  colorTo: purple
6
  sdk: gradio
 
8
  app_file: app.py
9
  pinned: false
10
  license: apache-2.0
11
+ short_description: Fast 8-step monocular novel view synthesis
12
  python_version: "3.10"
13
  startup_duration_timeout: 1h
14
  ---
15
 
16
+ # MetaView ⚡ Lightning — Monocular Novel View Synthesis
17
 
18
  Interactive demo for [**MetaView**](https://huggingface.co/Kwai-Kolors/MetaView)
19
  (ECCV 2026): *Monocular Novel View Synthesis with Scale-Aware Implicit Geometry
20
+ Priors* — accelerated with the **Qwen-Image-Edit Lightning 8-step LoRA**.
21
 
22
  Upload a single image, choose a target camera **yaw** (left/right) and **pitch**
23
+ (up/down), and MetaView renders the scene from the new viewpoint in **8 steps**
24
+ (down from 40), with no classifier-free guidance.
25
 
26
  The pipeline combines:
27
  - **Qwen-Image-Edit** MM-DiT backbone (novel-view generation),
28
+ - **[Qwen-Image-Edit Lightning](https://huggingface.co/lightx2v/Qwen-Image-Lightning)**
29
+ 8-step distillation LoRA (rank 64 / alpha 8 → scale 0.125), fused into the DiT,
30
  - **Depth-Anything-3** GIANT (implicit 3D feature priors) + NESTED (dense depth),
31
  - **Modified RoPE (PRoPE)** for metric scale anchoring of the camera pose.
32
 
33
+ The repeated DiT block is AoTI-compiled (LoRA-fused) offline by
34
+ [`hugging-apps/metaview-lightning-aoti-compile`](https://huggingface.co/spaces/hugging-apps/metaview-lightning-aoti-compile)
35
+ and served from the [`hugging-apps/metaview-lightning-aoti`](https://huggingface.co/datasets/hugging-apps/metaview-lightning-aoti)
36
+ dataset.
37
+
38
  Links: [Model](https://huggingface.co/Kwai-Kolors/MetaView) ·
39
  [Code](https://github.com/KlingAIResearch/MetaView) ·
40
  [Paper](https://arxiv.org/abs/2607.12000)