Spaces:

---

FrameAI4687

/

Omni-Video-Factory

like 1.3k

Running on Zero

Hi FrameAI team,

First of all – huge respect for what you built. I’ve been digging into your Omni-Image-Editor and Omni-Video-Factory Spaces, and I reverse-engineered the app.pyc – the architecture is beautiful: the 8B unified DiT (OmniMMDitV2), RoPE + SwiGLU, and especially the PiFlowPolicyNetwork / OmniScheduler for multi-step velocity prediction. It’s exactly the kind of unified image+video approach I’ve been prototyping.

I understand the production checkpoint (“OmniVideo”) is private and loaded from your internal repo – the public Spaces only ship the UI wrapper, which is why a clone runs on CPU but can’t generate.

Would you consider granting research-only access to the Omni Creator 2.0 weights? I’m working on:

comparing your hybrid flow-diffusion scheduler against standard DDIM
building a lightweight demo that credits you fully

I’m happy to sign any non-distribution agreement, keep the weights private, and share back any benchmarks or optimizations I find on low-VRAM setups.

Thanks for open-sourcing the Spaces and for the inspiration – even without weights, the design taught me a lot.

Best,
desirel