whatmodel.sh / video / Pyramid Flow (SD3)

Pyramid Flow (SD3)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Pyramid Flow (SD3) is an open-weights video model from Peking University / Kuaishou, released October 2024. Its published size is 2B parameters. Licence: stabilityai-ai-community.

Model information sources

· Parameter count ↗

2B = the SD3-Medium-based MM-DiT (24 layers) as the makers state it; the T5 and CLIP text encoders and the 3D VAE are not included.

42.0±3.2#67 of 93 video modelsfrom 1 benchmark result across 1 source* thinly tested, treat the rank as provisional

$ how it compares

The video models ranked either side of Pyramid Flow (SD3), out of 139 with a score.

#modelscorewhereRAMreleased
94Ray 2 Luma AI43±2comparison only2025-01
95Open-Sora 2.0 HPC-AI Tech42±5local~55GB2025-03
96Pyramid Flow (SD3) Peking University / Kuaishou42±3local~8GB2024-10
97Luma Dream Machine 1.0 Luma AI42±7comparison only2024-06
98Sora (original, 2024) OpenAI42±2comparison only2024-12

see all video models · find the ones that fit your machine

$ what it's good at

text to video▆▄▃▄▆▅█▃▂▂▂▄ #43 of 68 * · 96% of its score · 1 test

Beats Open-Sora 2.0 here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
GenAI-Arena text-to-video (anonymous Elo)1113.9 ✎▃▃···▅▃·▃▅·█ beats 8/10avg 1229▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 96%sharp

$ where to run it

Other ways to run it

  1. Run it as published · needs ~8 GB, NVIDIA only · from github.com/jy0205
    git clone https://github.com/jy0205/Pyramid-Flow && cd Pyramid-Flow && pip install -r requirements.txt && python -c "import torch;from huggingface_hub import snapshot_download;from pyramid_dit import PyramidDiTForVideoGeneration;from diffusers.utils import export_to_video;snapshot_download('rain1011/pyramid-flow-sd3',local_dir='pyramid-flow-sd3');m=PyramidDiTForVideoGeneration('pyramid-flow-sd3',model_name='pyramid_mmdit',model_dtype='bf16',model_variant='diffusion_transformer_768p');m.vae.enable_tiling();m.enable_sequential_cpu_offload();torch.set_grad_enabled(False);torch.autocast('cuda',dtype=torch.bfloat16).__enter__();export_to_video(m.generate(prompt='A movie trailer featuring the adventures of the 30 year old space man wearing a red wool knitted motorcycle helmet, blue sky, salt desert, cinematic style',num_inference_steps=[20,20,20],video_num_inference_steps=[10,10,10],height=768,width=1280,temp=16,guidance_scale=9.0,video_guidance_scale=5.0,output_type='pil',save_memory=True),'pyramid.mp4',fps=24)"

Weights and documentation: huggingface.co/rain1011 ↗

$ how the score was worked out · GenAI-Arena (TIGER-Lab) text-… 96%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
GenAI-Arena (TIGER-Lab) text-to-video leaderboard1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 96%MIT
prior · pull toward the middle▇ 4%
weighted average= 42.0

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json