base_model:
- rain1011/pyramid-flow-sd3
pipeline_tag: text-to-video
library_name: diffusers
Converted to bfloat16 from rain1011/pyramid-flow-sd3. Use the text encoders and tokenizers from that repo (or from SD3), no point reuploading them over and over unchanged.
Inference code is available here: github.com/jy0205/Pyramid-Flow.
Both 384p and 768p work on 24 GB VRAM. For 16 steps (5 second video), 384p takes a little over a minute on a 3090, and 768p takes about 7 minutes. For 31 steps (10 second video), 384p took about 10 minutes.
In diffusion_schedulers/scheduling_flow_matching.py, in the function init_sigmas_for_each_stage, one small change needs to be made:
Change this line:
self.timesteps_per_stage[i_s] = torch.from_numpy(timesteps[:-1])
To this:
self.timesteps_per_stage[i_s] = timesteps[:-1]
This will allow the model to be compatible with newer versions of pytorch and other libraries than is shown in the requirements.
Working with torch2.4.1+cu124.