code README: GitHub is the code home; download paths
Browse files- code/README.md +6 -5
code/README.md
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
|
| 3 |
Wizard is a real-time playable world model built on [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B). A lightweight engine keeps the world state and renders a *proxy video* (character skeletons + scene geometry); the video model in this repository turns the proxy, a first frame, a text prompt, up to two character reference images and a few retrieved memory frames into the picture.
|
| 4 |
|
| 5 |
-
This directory holds the model, the four training stages and offline inference. The
|
| 6 |
|
| 7 |
## Method in one page
|
| 8 |
|
|
@@ -67,18 +67,19 @@ Tested with Python 3.12, PyTorch 2.11 (CUDA 12.8) on 8 × H200. One 5B checkpoin
|
|
| 67 |
|
| 68 |
## Inference
|
| 69 |
|
| 70 |
-
A *case* is a directory produced by the engine: `first_frame.png`, `proxy.mkv` (proxy video), `cond.mp4` (the same proxy in the earlier colour coding; only the character layer is read from it), `tier.npy` (geometry tier per cell), `gbuf.npz` (depth and cameras), `slots.npz` (character regions), `prompt.txt`.
|
| 71 |
|
| 72 |
```bash
|
| 73 |
# bidirectional model: 30 steps, guidance 3
|
| 74 |
-
|
|
|
|
| 75 |
INF_UNK_NOCOORD=1 python -m wizard.inference.bidirectional \
|
| 76 |
-
--dit
|
| 77 |
--ref1 $C/ref_a.png --ref2 $C/ref_b.png --ref-slots 2,3 --out out.mp4
|
| 78 |
|
| 79 |
# causal 2-step student
|
| 80 |
INF_UNK_NOCOORD=1 python -m wizard.inference.causal \
|
| 81 |
-
--dit
|
| 82 |
--ref1 $C/ref_a.png --ref2 $C/ref_b.png --ref-slots 2,3 \
|
| 83 |
--cm_steps 1000,500 --cfg 1 --window 20 --sink 1 --out out_causal.mp4
|
| 84 |
```
|
|
|
|
| 2 |
|
| 3 |
Wizard is a real-time playable world model built on [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B). A lightweight engine keeps the world state and renders a *proxy video* (character skeletons + scene geometry); the video model in this repository turns the proxy, a first frame, a text prompt, up to two character reference images and a few retrieved memory frames into the picture.
|
| 4 |
|
| 5 |
+
This directory holds the model, the four training stages and offline inference. The code lives on GitHub at [wyzardWu/wizard](https://github.com/wyzardWu/wizard) (`code/`) and is mirrored in the `code/` folder of [`yuzhewu207/wizard`](https://huggingface.co/datasets/yuzhewu207/wizard), which also holds the raw data, the weights (`weights/`) and ready-made inference cases (`examples/`). The pre-encoded training data is in [`yuzhewu207/wizard-precoded`](https://huggingface.co/datasets/yuzhewu207/wizard-precoded).
|
| 6 |
|
| 7 |
## Method in one page
|
| 8 |
|
|
|
|
| 67 |
|
| 68 |
## Inference
|
| 69 |
|
| 70 |
+
A *case* is a directory produced by the engine: `first_frame.png`, `proxy.mkv` (proxy video), `cond.mp4` (the same proxy in the earlier colour coding; only the character layer is read from it), `tier.npy` (geometry tier per cell), `gbuf.npz` (depth and cameras), `slots.npz` (character regions), `prompt.txt`. Ready-made cases with our outputs are in the `examples/` folder of [`yuzhewu207/wizard`](https://huggingface.co/datasets/yuzhewu207/wizard).
|
| 71 |
|
| 72 |
```bash
|
| 73 |
# bidirectional model: 30 steps, guidance 3
|
| 74 |
+
huggingface-cli download yuzhewu207/wizard --repo-type dataset --include "weights/bidirectional/*" "weights/causal_dmd/*" "examples/*" --local-dir hf_wizard
|
| 75 |
+
C=hf_wizard/examples/fight_character_strikes_player
|
| 76 |
INF_UNK_NOCOORD=1 python -m wizard.inference.bidirectional \
|
| 77 |
+
--dit hf_wizard/weights/bidirectional/dit_step35000.safetensors --case $C --frames 121 \
|
| 78 |
--ref1 $C/ref_a.png --ref2 $C/ref_b.png --ref-slots 2,3 --out out.mp4
|
| 79 |
|
| 80 |
# causal 2-step student
|
| 81 |
INF_UNK_NOCOORD=1 python -m wizard.inference.causal \
|
| 82 |
+
--dit hf_wizard/weights/causal_dmd/ema_step480.safetensors --case $C --frames 121 \
|
| 83 |
--ref1 $C/ref_a.png --ref2 $C/ref_b.png --ref-slots 2,3 \
|
| 84 |
--cm_steps 1000,500 --cfg 1 --window 20 --sink 1 --out out_causal.mp4
|
| 85 |
```
|