wan2.2

.gitignore

@@ -233,3 +233,7 @@ time_cost.json
 costtime.json
 batch_eval.txt
 *.log
+
+# README comparison assets intentionally included in the PR
+!assets/wan21_comparison/
+!assets/wan21_comparison/*.mp4

NaviCache4Wan2.1/README.md

@@ -1,10 +1,10 @@
 # NaviCache for Wan2.1
 
-NaviCache is a training-free test-time self-calibration caching method for accelerating video diffusion models. This directory provides the NaviCache generation script for [Wan2.1](https://github.com/Wan-Video/Wan2.1).
+NaviCache is a training-free, test-time self-calibration caching method for accelerating [Wan2.1](https://github.com/Wan-Video/Wan2.1) text-to-video and image-to-video generation.
 
 ## Usage
 
-Follow [Wan2.1](https://github.com/Wan-Video/Wan2.1) to clone the repo, finish the installation, and download the required model weights. Then copy `navicache_generate.py` in this repo to the Wan2.1 repo.
+Install Wan2.1 following the [official repository](https://github.com/Wan-Video/Wan2.1), then copy the NaviCache integration script into the Wan2.1 repository:
 
 ```bash
 git clone https://github.com/Wan-Video/Wan2.1.git
@@ -14,11 +14,7 @@ cp NaviCache/NaviCache4Wan2.1/navicache_generate.py Wan2.1/
 cd Wan2.1
 ```
 
-Please make sure that `--ckpt_dir` points to the matching Wan2.1 checkpoint directory. The T2V-14B and T2V-1.3B models use different weights. The I2V 480P and I2V 720P models also use different weights. 
-
-## Text-to-Video
-
-For T2V with the 1.3B model:
+### T2V 1.3B, 832x480
 
 ```bash
 python navicache_generate.py \
@@ -27,12 +23,13 @@ python navicache_generate.py \
     --ckpt_dir ./Wan2.1-T2V-1.3B \
     --offload_model True \
     --t5_cpu \
+    --base_seed 42 \
     --navicache_thresh 0.05 \
     --navicache_align_steps 10 \
-    --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
+    --prompt "A boat sailing leisurely along the Seine River with the Eiffel Tower in background, zoom out"
 ```
 
-For T2V with the 14B model:
+### T2V 14B, 1280x720
 
 ```bash
 python navicache_generate.py \
@@ -41,14 +38,13 @@ python navicache_generate.py \
     --ckpt_dir ./Wan2.1-T2V-14B \
     --offload_model True \
     --t5_cpu \
+    --base_seed 42 \
     --navicache_thresh 0.05 \
     --navicache_align_steps 10 \
-    --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
+    --prompt "A person is skateboarding"
 ```
 
-## Image-to-Video
-
-For I2V with the 480P model:
+### I2V 14B, 832x480
 
 ```bash
 python navicache_generate.py \
@@ -58,28 +54,110 @@ python navicache_generate.py \
     --image examples/i2v_input.JPG \
     --offload_model True \
     --t5_cpu \
+    --base_seed 42 \
     --navicache_thresh 0.05 \
     --navicache_align_steps 10 \
-    --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds."
+    --prompt "A beautiful coastal beach in spring, waves lapping on sand by Hokusai, in the style of Ukiyo"
 ```
 
-For I2V with the 720P model:
+### I2V 14B, 1280x720
 
 ```bash
 python navicache_generate.py \
     --task i2v-14B \
     --size 1280*720 \
     --ckpt_dir ./Wan2.1-I2V-14B-720P \
     --image examples/i2v_input.JPG \
+    --frame_num 61 \
     --offload_model True \
     --t5_cpu \
+    --base_seed 42 \
     --navicache_thresh 0.05 \
     --navicache_align_steps 10 \
-    --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds."
+    --prompt "an orange cat"
 ```
 
-The I2V commands use the sample image included in Wan2.1 at `examples/i2v_input.JPG`; replace it with your own input image as needed.Generated files are written to `outputs/` unless `--save_file` is specified. Generation timing records are appended to `output/time_cost.json` by default. Use `--out_dir` to choose another timing-output directory.
+## Results
+
+### Inference Latency Comparison
+
+| Wan2.1 configuration | Resolution | Wan2.1 | TeaCache | TeaCache speedup | NaviCache | NaviCache speedup |
+|:---------------------|:----------:|:-------------------:|:----------------------------:|:----------------:|:----------------------:|:-----------------:|
+| T2V 1.3B | 832x480 | 488.0 s | 209.5 s | 2.33x | 160.1 s | 3.05x |
+| T2V 14B | 1280x720 | 3528.1 s | 1784.2 s | 1.98x | 1361.1 s | 2.59x |
+| I2V 14B 480P | 832x480 | 825.9 s | 604.4 s | 1.37x | 481.4 s | 1.72x |
+| I2V 14B 720P | 1280x720 | 1887.2 s | 1296.2 s | 1.46x | 1007.9 s | 1.87x |
+
+### Visual Quality Comparison
+
+#### T2V 1.3B, 832x480
+
+<div align="center">
+  <video src="../assets/wan21_comparison/t2v13b480_boat_comparison.mp4" width="936" controls muted loop></video>
+</div>
+
+| Wan2.1 | TeaCache | NaviCache |
+|:------:|:--------:|:---------:|
+| 488.0 s (1.00x) | 209.5 s (2.33x) | **160.1 s (3.05x)** |
+
+<details>
+<summary>Prompt: A boat sailing leisurely along the Seine River with the Eiffel Tower in background, zoom out</summary>
+
+A boat sailing leisurely along the Seine River with the Eiffel Tower in background, zoom out
+
+</details>
+
+#### T2V 14B, 1280x720
+
+<div align="center">
+  <video src="../assets/wan21_comparison/t2v14b720_skateboarding_comparison.mp4" width="960" controls muted loop></video>
+</div>
+
+| Wan2.1 | TeaCache | NaviCache |
+|:------:|:--------:|:---------:|
+| 3528.1 s (1.00x) | 1784.2 s (1.98x) | **1361.1 s (2.59x)** |
+
+<details>
+<summary>Prompt: A person is skateboarding</summary>
+
+A person is skateboarding
+
+</details>
+
+#### I2V 14B, 832x480
+
+<div align="center">
+  <video src="../assets/wan21_comparison/i2v14b480_ribbon_dancer_comparison.mp4" width="936" controls muted loop></video>
+</div>
+
+| Wan2.1 | TeaCache | NaviCache |
+|:------:|:--------:|:---------:|
+| 825.9 s (1.00x) | 604.4 s (1.37x) | **481.4 s (1.72x)** |
+
+<details>
+<summary>Prompt</summary>
+
+The ribbon dancer spins rapidly across the courtyard, both red silk ribbons tracing wide fluid arcs as her layered costume billows; the camera tracks sideways with strong natural background parallax.
+
+</details>
+
+#### I2V 14B, 1280x720
+
+<div align="center">
+  <video src="../assets/wan21_comparison/i2v14b720_clockwork_hummingbird_comparison.mp4" width="960" controls muted loop></video>
+</div>
+
+| Wan2.1 | TeaCache | NaviCache |
+|:------:|:--------:|:---------:|
+| 1887.2 s (1.00x) | 1296.2 s (1.46x) | **1007.9 s (1.87x)** |
+
+<details>
+<summary>Prompt</summary>
+
+The clockwork hummingbird beats both articulated wings rapidly and flies from one red flower to the next; tiny gears turn visibly while the camera arcs around it through the greenhouse with deep parallax.
+
+</details>
 
 ## Acknowledgements
 
-We would like to thank the contributors to [Wan2.1](https://github.com/Wan-Video/Wan2.1).
+We would like to thank the contributors to [Wan2.1](https://github.com/Wan-Video/Wan2.1), [TeaCache](https://github.com/ali-vilab/TeaCache), [EasyCache](https://github.com/H-EmbodVis/EasyCache), and [MagCache](https://github.com/Zehong-Ma/MagCache).

NaviCache4Wan2.2/README.md

@@ -0,0 +1,66 @@
+# NaviCache for Wan2.2
+
+NaviCache is a training-free test-time self-calibration caching method for accelerating video diffusion models. This directory provides the NaviCache generation script for [Wan2.2](https://github.com/Wan-Video/Wan2.2).
+
+## Usage
+
+Follow the official [Wan2.2](https://github.com/Wan-Video/Wan2.2) instructions to clone the repository, install the required environment, and download the Wan2.2-TI2V-5B model weights. NaviCache requires no training or offline calibration. From the directory containing the cloned `Wan2.2` repository, copy `navicache_generate.py` to its root:
+
+```bash
+git clone https://github.com/HelloZicky/NaviCache.git
+
+cp NaviCache/NaviCache4Wan2.2/navicache_generate.py Wan2.2/
+cd Wan2.2
+mkdir -p outputs
+```
+
+Make sure that `--ckpt_dir` points to the downloaded Wan2.2-TI2V-5B checkpoint directory. Once the official environment and weights are ready, run either of the following commands to generate videos with NaviCache.
+
+## Text-to-Video
+
+Run Wan2.2-TI2V-5B without an input image:
+
+```bash
+python navicache_generate.py \
+    --task ti2v-5B \
+    --size 1280*704 \
+    --frame_num 121 \
+    --ckpt_dir ./Wan2.2-TI2V-5B \
+    --offload_model True \
+    --convert_model_dtype \
+    --t5_cpu \
+    --navicache_thresh 0.05 \
+    --navicache_align_steps 10 \
+    --navicache_process_noise 0.05 \
+    --navicache_measurement_noise 0.05 \
+    --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
+    --save_file ./outputs/wan22_t2v_navicache.mp4
+```
+
+## Image-to-Video
+
+Pass an input image to the same TI2V checkpoint with `--image`:
+
+```bash
+python navicache_generate.py \
+    --task ti2v-5B \
+    --size 1280*704 \
+    --frame_num 121 \
+    --ckpt_dir ./Wan2.2-TI2V-5B \
+    --image examples/i2v_input.JPG \
+    --offload_model True \
+    --convert_model_dtype \
+    --t5_cpu \
+    --navicache_thresh 0.05 \
+    --navicache_align_steps 10 \
+    --navicache_process_noise 0.05 \
+    --navicache_measurement_noise 0.05 \
+    --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard at a sunny beach." \
+    --save_file ./outputs/wan22_i2v_navicache.mp4
+```
+
+The command uses the sample image included in Wan2.2 at `examples/i2v_input.JPG`; replace it with your own input image as needed. Generated videos are written to the path provided by `--save_file`.
+
+## Acknowledgements
+
+We would like to thank the contributors to [Wan2.2](https://github.com/Wan-Video/Wan2.2).