hudabey commited on
Commit
a39b100
·
verified ·
1 Parent(s): 41fe819

Rename to StrideAR

Browse files
Files changed (1) hide show
  1. README.md +27 -27
README.md CHANGED
@@ -22,19 +22,19 @@ base_model:
22
  - BestWishYsh/Helios-Distilled
23
  ---
24
 
25
- # SpecAR: Distributed 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
- SpecAR is a training-free inference method for accelerating autoregressive video generation. Instead of invoking a large target model for every video block, SpecAR 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
- SpecAR 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/) · [Paper](https://statite.ai/specar/specar.pdf) · [GitHub](https://github.com/statite-ai/SpecAR) · [ComfyUI](https://github.com/statite-ai/ComfyUI-SpecAR)**
36
 
37
- ![SpecAR speed comparison on Krea Realtime 14B](https://raw.githubusercontent.com/statite-ai/SpecAR/main/assets/race.webp)
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
 
@@ -42,7 +42,7 @@ SpecAR supports **text-to-video, image-to-video, and video-to-video generation**
42
 
43
  ### H100: Published evaluation
44
 
45
- SpecAR 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
  |---|---:|---|
@@ -55,19 +55,19 @@ For Krea Realtime 14B, generating a 26.6-second video:
55
  |---|---:|---:|
56
  | Target-only | 87.6 s | 1.00× |
57
  | SDVG, operating point for k=2 | 56.8 s | 1.54× |
58
- | **SpecAR, k=2** | **32.8 s** | **2.67×** |
59
- | SpecAR, k=5 | 39.0 s | 2.25× |
60
- | SpecAR, 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, SpecAR 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, SpecAR 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 SpecAR 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
 
@@ -82,7 +82,7 @@ The B200 run completed the 26.6-second video in less than its playback duration.
82
 
83
  ## Supported models
84
 
85
- SpecAR uses existing publicly released model weights. **No training or new checkpoint is required.**
86
 
87
  | Backbone | Target | Drafter |
88
  |---|---|---|
@@ -97,7 +97,7 @@ SpecAR uses existing publicly released model weights. **No training or new check
97
  | H200 | Supported | Supported | FA3 |
98
  | B200 | Supported | Not yet supported | FA4 |
99
 
100
- SpecAR 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
 
@@ -118,11 +118,11 @@ specar download --backbone krea
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. SpecAR 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 SpecAR
126
 
127
  ### 1. Text-to-video
128
 
@@ -135,7 +135,7 @@ specar generate --backbone krea \
135
  --out reef.mp4
136
  ```
137
 
138
- By default, SpecAR generates 36 blocks, producing approximately 26.6 seconds of video at 16 FPS.
139
 
140
  ### 2. Image-to-video
141
 
@@ -183,21 +183,21 @@ The `--strength` parameter controls the degree of transformation relative to the
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 [SpecAR GitHub repository](https://github.com/statite-ai/SpecAR).
187
 
188
  ## ComfyUI
189
 
190
- SpecAR 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 SpecAR 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
- SpecAR 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
 
@@ -214,7 +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, SpecAR requires:
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
- SpecAR is an approximate decoder: the scheduling guarantee does not guarantee identical target-model outputs or preservation of the target distribution.
227
 
228
  ## Reproducibility
229
 
@@ -242,8 +242,8 @@ See the [repository documentation](https://github.com/statite-ai/SpecAR) for pin
242
  ## Citation
243
 
244
  ```bibtex
245
- @misc{hassan2026specar,
246
- title = {SpecAR: Distributed Target Blocks for Fast Autoregressive Video Generation},
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
- SpecAR'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
 
@@ -265,4 +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/) · [Paper](https://statite.ai/specar/specar.pdf) · [Source Code](https://github.com/statite-ai/SpecAR) · [ComfyUI](https://github.com/statite-ai/ComfyUI-SpecAR)
 
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
+ ![StrideAR speed comparison on Krea Realtime 14B](https://raw.githubusercontent.com/statite-ai/SpecAR/main/assets/race.webp)
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)