Rename to StrideAR
Browse files
README.md
CHANGED
|
@@ -22,19 +22,19 @@ base_model:
|
|
| 22 |
- BestWishYsh/Helios-Distilled
|
| 23 |
---
|
| 24 |
|
| 25 |
-
#
|
| 26 |
|
| 27 |
**Up to 10.54× faster autoregressive video generation, without retraining or auxiliary reward models.**
|
| 28 |
|
| 29 |
-
|
| 30 |
|
| 31 |
This eliminates online perceptual routing and its associated scoring overhead, while controlling the longest uninterrupted sequence of drafted blocks.
|
| 32 |
|
| 33 |
-
|
| 34 |
|
| 35 |
-
**[Project Page](https://statite.ai/stridear/) · [
|
| 36 |
|
| 37 |
-
**
|
| 193 |
|
| 194 |
-
The ComfyUI integration uses the same
|
| 195 |
|
| 196 |
See the custom-node repository for its installation requirements and supported workflows.
|
| 197 |
|
| 198 |
## How it works
|
| 199 |
|
| 200 |
-
|
| 201 |
|
| 202 |
Given a video of H autoregressive blocks and a target budget k, it selects target positions according to
|
| 203 |
|
|
@@ -214,7 +214,7 @@ The target generates the blocks at these positions. The drafter generates all re
|
|
| 214 |
|
| 215 |
The resulting schedule minimizes the longest uninterrupted draft run for the specified target budget, with the first block assigned to the target.
|
| 216 |
|
| 217 |
-
Unlike online perceptual routing,
|
| 218 |
|
| 219 |
- No additional training.
|
| 220 |
- No learned router or reward model.
|
|
@@ -223,7 +223,7 @@ Unlike online perceptual routing, SpecAR requires:
|
|
| 223 |
|
| 224 |
Both models operate on the same committed video trajectory. Model-specific KV caches are reconstructed from the committed context as required.
|
| 225 |
|
| 226 |
-
|
| 227 |
|
| 228 |
## Reproducibility
|
| 229 |
|
|
@@ -242,8 +242,8 @@ See the [repository documentation](https://github.com/statite-ai/SpecAR) for pin
|
|
| 242 |
## Citation
|
| 243 |
|
| 244 |
```bibtex
|
| 245 |
-
@misc{
|
| 246 |
-
title = {
|
| 247 |
author = {Hassan, Hudeifa},
|
| 248 |
year = {2026},
|
| 249 |
note = {Preprint},
|
|
@@ -253,7 +253,7 @@ See the [repository documentation](https://github.com/statite-ai/SpecAR) for pin
|
|
| 253 |
|
| 254 |
## License
|
| 255 |
|
| 256 |
-
|
| 257 |
|
| 258 |
The Krea Realtime backend incorporates code adapted from Krea's inference implementation and retains **CC BY-NC-SA 4.0** licensing, restricting commercial use. Vendored Helios components retain their respective Apache-2.0 terms.
|
| 259 |
|
|
@@ -265,4 +265,4 @@ See [LICENSE and file-level notices](https://github.com/statite-ai/SpecAR#licens
|
|
| 265 |
|
| 266 |
Developed by **[Statite](https://statite.ai)**.
|
| 267 |
|
| 268 |
-
[Project Page](https://statite.ai/stridear/) · [
|
|
|
|
| 22 |
- BestWishYsh/Helios-Distilled
|
| 23 |
---
|
| 24 |
|
| 25 |
+
# StrideAR: Strided Target Blocks for Fast Autoregressive Video Generation
|
| 26 |
|
| 27 |
**Up to 10.54× faster autoregressive video generation, without retraining or auxiliary reward models.**
|
| 28 |
|
| 29 |
+
StrideAR is a training-free inference method for accelerating autoregressive video generation. Instead of invoking a large target model for every video block, StrideAR assigns most blocks to a smaller drafter and distributes a limited number of target-generated blocks evenly across the video.
|
| 30 |
|
| 31 |
This eliminates online perceptual routing and its associated scoring overhead, while controlling the longest uninterrupted sequence of drafted blocks.
|
| 32 |
|
| 33 |
+
StrideAR supports **text-to-video, image-to-video, and video-to-video generation** on Krea Realtime 14B, with a single implementation supporting NVIDIA Hopper and Blackwell GPUs.
|
| 34 |
|
| 35 |
+
**[Project Page](https://statite.ai/stridear/) · [Earlier preprint](https://statite.ai/specar/specar.pdf) · [GitHub](https://github.com/statite-ai/SpecAR) · [ComfyUI](https://github.com/statite-ai/ComfyUI-SpecAR)**
|
| 36 |
|
| 37 |
+

|
| 38 |
|
| 39 |
*Generation comparison on Krea Realtime 14B using one NVIDIA H100. The methods generate the same video duration from the same prompt and seed. Reported latencies come from the paper's evaluation.*
|
| 40 |
|
|
|
|
| 42 |
|
| 43 |
### H100: Published evaluation
|
| 44 |
|
| 45 |
+
StrideAR was evaluated on Krea Realtime 14B and Helios 14B using a single NVIDIA H100 80GB per model pair.
|
| 46 |
|
| 47 |
| Model | Speedup over target-only | GPU |
|
| 48 |
|---|---:|---|
|
|
|
|
| 55 |
|---|---:|---:|
|
| 56 |
| Target-only | 87.6 s | 1.00× |
|
| 57 |
| SDVG, operating point for k=2 | 56.8 s | 1.54× |
|
| 58 |
+
| **StrideAR, k=2** | **32.8 s** | **2.67×** |
|
| 59 |
+
| StrideAR, k=5 | 39.0 s | 2.25× |
|
| 60 |
+
| StrideAR, k=7 | 42.9 s | 2.04× |
|
| 61 |
|
| 62 |
These results use 36-block rollouts at 832×480 resolution, BF16 precision, five denoising steps and FlashAttention-3.
|
| 63 |
|
| 64 |
+
At k=7, StrideAR retains **99.7% of target-only Restricted VBench Quality-6**, while generating 2.04× faster than target-only inference.
|
| 65 |
|
| 66 |
+
Across the evaluated operating points on Krea Realtime and Helios, StrideAR also achieves lower latency and matching or higher measured VisionReward and Restricted VBench Quality-6 than SDVG.
|
| 67 |
|
| 68 |
### B200: Blackwell execution
|
| 69 |
|
| 70 |
+
The same StrideAR Krea implementation supports NVIDIA B200 GPUs through automatic FlashAttention-4 selection. No separate Blackwell implementation or model conversion is required.
|
| 71 |
|
| 72 |
In an engineering validation using a 26.6-second coral-reef generation at k=7:
|
| 73 |
|
|
|
|
| 82 |
|
| 83 |
## Supported models
|
| 84 |
|
| 85 |
+
StrideAR uses existing publicly released model weights. **No training or new checkpoint is required.**
|
| 86 |
|
| 87 |
| Backbone | Target | Drafter |
|
| 88 |
|---|---|---|
|
|
|
|
| 97 |
| H200 | Supported | Supported | FA3 |
|
| 98 |
| B200 | Supported | Not yet supported | FA4 |
|
| 99 |
|
| 100 |
+
StrideAR automatically selects a compatible attention backend on supported hardware. The inference algorithm, installation procedure and CLI remain the same across Hopper and Blackwell.
|
| 101 |
|
| 102 |
## Installation
|
| 103 |
|
|
|
|
| 118 |
|
| 119 |
The Krea setup downloads approximately 50 GB of public model files, including the Krea Realtime 14B target, Self-Forcing 1.3B drafter, text encoder and VAE.
|
| 120 |
|
| 121 |
+
The same installation works on H100 and B200. StrideAR automatically uses FA3 on Hopper and FA4 on Blackwell.
|
| 122 |
|
| 123 |
On shared GPU machines, running `hf auth login` before downloading can help avoid anonymous Hugging Face rate limits.
|
| 124 |
|
| 125 |
+
## Generate with StrideAR
|
| 126 |
|
| 127 |
### 1. Text-to-video
|
| 128 |
|
|
|
|
| 135 |
--out reef.mp4
|
| 136 |
```
|
| 137 |
|
| 138 |
+
By default, StrideAR generates 36 blocks, producing approximately 26.6 seconds of video at 16 FPS.
|
| 139 |
|
| 140 |
### 2. Image-to-video
|
| 141 |
|
|
|
|
| 183 |
|
| 184 |
The default target budget is k=7. Lower values generally favor speed, while increasing the target budget changes the quality–latency trade-off.
|
| 185 |
|
| 186 |
+
For all options and sample outputs, see the [StrideAR GitHub repository](https://github.com/statite-ai/SpecAR).
|
| 187 |
|
| 188 |
## ComfyUI
|
| 189 |
|
| 190 |
+
StrideAR is also available as a ComfyUI custom-node integration for Krea Realtime 14B.
|
| 191 |
|
| 192 |
**[ComfyUI-SpecAR — Installation, Nodes and Workflows](https://github.com/statite-ai/ComfyUI-SpecAR)**
|
| 193 |
|
| 194 |
+
The ComfyUI integration uses the same StrideAR inference engine rather than a separate implementation of the decoding algorithm.
|
| 195 |
|
| 196 |
See the custom-node repository for its installation requirements and supported workflows.
|
| 197 |
|
| 198 |
## How it works
|
| 199 |
|
| 200 |
+
StrideAR treats the placement of expensive target-model executions as a scheduling problem.
|
| 201 |
|
| 202 |
Given a video of H autoregressive blocks and a target budget k, it selects target positions according to
|
| 203 |
|
|
|
|
| 214 |
|
| 215 |
The resulting schedule minimizes the longest uninterrupted draft run for the specified target budget, with the first block assigned to the target.
|
| 216 |
|
| 217 |
+
Unlike online perceptual routing, StrideAR requires:
|
| 218 |
|
| 219 |
- No additional training.
|
| 220 |
- No learned router or reward model.
|
|
|
|
| 223 |
|
| 224 |
Both models operate on the same committed video trajectory. Model-specific KV caches are reconstructed from the committed context as required.
|
| 225 |
|
| 226 |
+
StrideAR is an approximate decoder: the scheduling guarantee does not guarantee identical target-model outputs or preservation of the target distribution.
|
| 227 |
|
| 228 |
## Reproducibility
|
| 229 |
|
|
|
|
| 242 |
## Citation
|
| 243 |
|
| 244 |
```bibtex
|
| 245 |
+
@misc{hassan2026stridear,
|
| 246 |
+
title = {StrideAR: Strided Target Blocks for Fast Autoregressive Video Generation},
|
| 247 |
author = {Hassan, Hudeifa},
|
| 248 |
year = {2026},
|
| 249 |
note = {Preprint},
|
|
|
|
| 253 |
|
| 254 |
## License
|
| 255 |
|
| 256 |
+
StrideAR's original scheduling, inference utilities, CLI and benchmarking code are distributed under **Apache-2.0**.
|
| 257 |
|
| 258 |
The Krea Realtime backend incorporates code adapted from Krea's inference implementation and retains **CC BY-NC-SA 4.0** licensing, restricting commercial use. Vendored Helios components retain their respective Apache-2.0 terms.
|
| 259 |
|
|
|
|
| 265 |
|
| 266 |
Developed by **[Statite](https://statite.ai)**.
|
| 267 |
|
| 268 |
+
[Project Page](https://statite.ai/stridear/) · [Earlier preprint](https://statite.ai/specar/specar.pdf) · [Source Code](https://github.com/statite-ai/SpecAR) · [ComfyUI](https://github.com/statite-ai/ComfyUI-SpecAR)
|