---
title: MiniMax-H3 360 Orbit Workflow
emoji: 🌀
colorFrom: yellow
colorTo: green
sdk: gradio
sdk_version: 6.29.0
python_version: '3.13'
app_file: app.py
pinned: false
hf_oauth: true
short_description: One photo into a 360 camera orbit workflow
models:
  - pablodawson/MiniMax-H3-360-Orbit-LoRA
  - multimodalart/MiniMax-H3-Pruned
  - MiniMaxAI/MiniMax-H3
---

# MiniMax-H3 360° Orbit — Gradio Workflow (ZeroGPU inline)

A Gradio Workflow (node-based canvas) that turns one photo into a 360° camera orbit video with the
scene held still — with the denoiser **inlined into this Space**: the orbit operator node is a bound
Python function (`orbit` in `orbit.py`, adapted from
[pablodawson/minimax-h3-360-orbit](https://huggingface.co/spaces/pablodawson/minimax-h3-360-orbit))
decorated with `@spaces.GPU`, so runs execute on **this Space's own ZeroGPU hardware**, not through a
gradio client call to the upstream Space.

The canvas wires a **Photo to orbit** reference (prefilled with an example) into the `orbit` function
node and out to two subjects: the **Orbit Video** and the **Run Report**. The five knobs — canvas,
steps, seed, LoRA strength, duration — are literal reference nodes prefilled with the LoRA card's
recommendations (**768x768**, **28 steps**, seed **904231**, strength **1.0**, **3 s** ≈ 73 frames).

One remote call remains by design: the 62 GiB Qwen3-VL **conditioner** half of the split MiniMax-H3
deployment still runs in
[multimodalart/qwen3vl-conditioner](https://huggingface.co/spaces/multimodalart/qwen3vl-conditioner)
and is called over the gradio API per request — MiniMax-H3's two halves cannot share one Space's
storage quota. Everything else (the pruned transformer partition, both VAEs, the orbit LoRA) loads
and runs here.

## Hardware

This Space needs **ZeroGPU** hardware (Settings → Hardware → ZeroGPU, requires PRO). On CPU hardware
the `@spaces.GPU` decorator is a no-op and generation will not work. Model load at startup pulls
~49 GB of weights, so first boot takes a while — watch the logs.

## Sign in to run

HF OAuth is enabled: sign in on the canvas to run the pipeline; the conditioner call is billed to
your identity. The Space owner can additionally edit the canvas and save back to `workflow.json`.