| # FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION |
|
|
| This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models. |
|
|
|
|
| ## 🚀 Installation |
|
|
| Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1. |
|
|
| --- |
|
|
| ## ▶️ Usage |
|
|
| ### 1. Single Video Generation |
|
|
| Run accelerated generation using FlowCache: |
|
|
| ```bash |
| # FlowCache for text-to-video generation |
| bash scripts/single_run/flowcache_t2v.sh |
| |
| # FlowCache for video-to-video generation |
| bash scripts/single_run/flowcache_v2v.sh |
| |
| # Baseline acceleration method (TeaCache) for text-to-video |
| bash scripts/single_run/teacache_t2v.sh |
| |
| # Baseline acceleration method (TeaCache) for video-to-video |
| bash scripts/single_run/teacache_v2v.sh |
| ``` |
|
|
| ### 2. Benchmark Sampling |
|
|
| Generate videos for evaluation on standard benchmarks: |
|
|
| ```bash |
| # VBench |
| bash scripts/sample/flowcache_vbench.sh |
| bash scripts/sample/teacache_vbench.sh |
| |
| # PhysicsIQ |
| bash scripts/sample/flowcache_physicsiq.sh |
| bash scripts/sample/teacache_physicsiq.sh |
| ``` |
|
|
| ### 3. Quality Evaluation |
|
|
| Compute perceptual and structural similarity metrics between original and accelerated generations: |
|
|
| ```bash |
| bash scripts/metric.sh |
| ``` |
|
|
| --- |
|
|
| ## ⚙️ Key Parameters |
|
|
| | Parameter | Description | |
| |----------|-------------| |
| | `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision | |
| | ` warmup_steps` | Number of denoising steps where reuse is disabled | |
| | `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained | |
| | `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` | |
| | `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` | |
| | `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) | |
| | `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) | |
| | `prompt` | Input prompt for conditional generation | |
| | `output_path` | Path to save generated videos | |
| | `config_file` | Path to MAGI-1 model configuration | |
|
|
| --- |
