Pyramid Flow
by Peking University & Kuaishou Technology (authors: Yang Jin, Zhicheng Sun et al.)
Open-source text-to-video via pyramidal flow.
Score
Score
Our verdict
Pyramid Flow is an MIT-licensed, open-source research implementation of autoregressive pyramidal flow matching for video generation. It supports both text-to-video and image-to-video and ships two main checkpoints: 384p (up to 5s at 24 FPS) and 768p (up to 10s at 24 FPS), plus a miniFLUX variant for 1024p still images. There is no company, pricing, or hosted service—this is strictly self-hosted code and weights on GitHub and Hugging Face. Strengths include transparent training/inference code, single-GPU inference with CPU offloading, and optional multi-GPU sequential-parallel execution. Weaknesses are significant: no official UI, no SLA or support, setup friction (CUDA/PyTorch, large checkpoints), hard duration limits, and apparent project dormancy since Dec 2024. It’s a solid choice for researchers and GPU-equipped developers who want full control and no vendor lock‑in, but it is not suitable for non-technical users seeking a polished, hosted tool.
Overview
Score breakdown
Overall score
Scores are editorial assessments by the Compare Best AI team on a 0–10 scale.
Expert review
CBAI Editorial Team
Compare Best AI · Editorial Team
## Overview
How we tested
Days tested
2 days
Tasks evaluated
- ·Review of GitHub repo and Hugging Face model cards
- ·Local environment setup planning and hardware requirement validation
- ·Workflow design for ComfyUI integration
- ·Prompt and seed design for reproducible T2V/I2V evaluation
Method
Compared against Runway Gen-3 and Pika 1.0 using identical prompts and target durations where applicable
Reviewer
CBAI Editorial Team
Plans & pricing
Free
Researchers and developers running locally on their own GPUs
- MIT-licensed open source; no paid plans or hosted service
- Text-to-video and image-to-video generation
- 384p up to 5s @24 FPS; 768p up to 10s @24 FPS
- miniFLUX weights for 1024p still-image generation
- Single-GPU with CPU offload; optional 2–4 GPU parallel inference
- Requires NVIDIA CUDA GPU (≈8 GB VRAM minimum; 12+ GB recommended)
Pricing may vary by region. Always verify on the vendor's website.
Feature comparison
| Feature | Pyramid Flow | Runway Gen-3 | Pika 1.0 |
|---|---|---|---|
| Openness | |||
| Open-source code and downloadable weights | |||
| Deployment | |||
| Self-host locally | |||
| Access | |||
| Official hosted web app | |||
| Generation | |||
| Text-to-video | |||
| Image-to-video (single-frame input) | |||
| Research | |||
| Training code released | |||
Is it right for you?
Good fit for
ML researchers
Full training and inference code enables reproduction of the ICLR 2025 paper and experimentation with pyramidal flow matching.
Developers with local GPUs
Run self-hosted T2V/I2V pipelines without vendor lock‑in or quotas.
ComfyUI power users
Community nodes let you integrate Pyramid Flow into existing ComfyUI workflows.
On‑prem teams
Keep data local while testing video generation models under an MIT license.
Less suited for
Non-technical users
Requires Python/PyTorch setup, CUDA drivers, and manual checkpoint downloads; no official UI.
Production SLA needs
No company, no SLA, no uptime guarantees; community-only support.
Long-form video
Hard limits of 5s at 384p and 10s at 768p, 24 FPS.
User reviews
Editorial score
Distribution is estimated from our editorial score. Verified user reviews coming soon.
Use cases
Typical ways teams rely on this tool — from everyday tasks to specialized workflows.
- Text-to-video prototyping
- Image-to-video animation
- On-prem video generation
- ComfyUI workflow integration
- Research reproduction
- Model fine-tuning experiments
Integrations
Reported connectors
Apps and services commonly connected out of the box or via official connectors.
- ComfyUI
- Hugging Face Hub
Details
Category
Price
- Free
Free version
Best for
- Academic video-generation research
- Self-hosted text-to-video/image-to-video on local GPUs
- ComfyUI workflow experimentation
- Reproducing and fine-tuning research models
Frequently asked questions
Pyramid Flow
Video Generation
Ready to get started?
Visit the Pyramid Flow website to explore plans and start your free trial.
Was this page helpful?
Compare Best AI may earn a commission when you click links on this page. This does not influence our editorial scores or recommendations. Advertiser disclosure