SpecAR

ComfyUI-SpecAR

Runs SpecAR text-to-video on Krea Realtime 14B in ComfyUI. It adds one node, SpecAR Generate, and calls the installed specar package.

Project page: statite.ai/specar · Paper: PDF · Code: statite-ai/SpecAR

Install

git clone https://github.com/comfyanonymous/ComfyUI && cd ComfyUI
python3.11 -m venv .venv-krea && source .venv-krea/bin/activate
pip install -r requirements.txt "specar[krea] @ git+https://github.com/statite-ai/SpecAR"
git clone https://github.com/statite-ai/ComfyUI-SpecAR custom_nodes/ComfyUI-SpecAR
specar download --backbone krea --weights models/specar/krea
python main.py

specar pins exact library versions, including torch, so use a separate environment. The download is about 50 GB.

Use

1. Open http://127.0.0.1:8188. If ComfyUI runs on a remote GPU machine, first run ssh -L 8188:127.0.0.1:8188 user@your-gpu-machine on your own computer.
2. Go to Templates, find ComfyUI-SpecAR and open SpecAR. You can also drag example_workflows/SpecAR.json onto the page.
3. Type your prompt in SpecAR Generate.

4. Click Run. The first run loads the models, which takes about 3 minutes. Later runs start right away.
5. The video plays in Save Video and is saved in output/video.

Settings

| Input | Meaning | |---|---| | prompt | The text prompt | | k (quality) | How much of the video the large target model generates; the small drafter generates the rest. Higher is slower, and 2 is the fastest setting | | seconds | Video length, rounded to the nearest 0.75 s. 26.6 s (426 frames) is the tested length | | seed | Random seed. Set it to fixed to reproduce a video | | weights | Weights folder; empty means models/specar/krea. You can change it without restarting ComfyUI | | keep_loaded | Keep the models loaded for the next run. Changing the prompt then takes under a second | | low_vram_offload | Keep only the active model on the GPU. On by default on GPUs under 75 GiB |

GPU memory

Measured on one H100 for a 26.6 s video:

| Setting | Peak GPU memory | Generation time, k = 2 | Generation time, k = 7 | |---|---|---|---| | low_vram_offload off | 44.5 GiB | 30 s | 39 s | | low_vram_offload on | 41.9 GiB | 34 s | 46 s |

Offload also ran within a 48 GB limit set on the H100. Either setting needs more than 40 GB of GPU memory, because the target model and its context cache must be on the GPU together. Offload keeps the model weights in system RAM, which we estimate at about 40 GB.

Tested on H100 (FlashAttention 3) and B200 (FlashAttention 4), which are selected automatically. Peak memory is the same on both, and on a B200 a 26.6 s video at k = 2 generates in 19 s. Other NVIDIA GPUs with enough memory need FlashAttention 2 (pip install flash-attn) and are untested. A different attention kernel can give a different video for the same seed, so the reference hashes apply only to FlashAttention 3 on Hopper GPUs.

Reproducibility

With FlashAttention 3 on an H100, the latents and 8-bit frames from this node matched specar generate bit for bit, and the latents matched the specar-benchmark reference hashes, with offload on or off. Save Video uses its own encoder, so the mp4 file differs from the one specar generate writes, and encoding adds about 30 s at 426 frames.

Citation

@misc{hassan2026specar,
  title  = {SpecAR: Distributed Target Blocks for Fast Autoregressive Video Generation},
  author = {Hassan, Hudeifa},
  year   = {2026},
  note   = {Preprint},
  url    = {https://statite.ai/specar}
}

License

This node is Apache-2.0. It imports specar and contains none of its code. SpecAR's Krea backend is CC BY-NC-SA 4.0, which allows non-commercial use only, and model weights carry their own licenses. See the SpecAR license notes.