Spaces:
Running on Zero
Running on Zero
| title: Video Replace Background workflow | |
| emoji: 🎬 | |
| colorFrom: indigo | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.29.1 | |
| app_file: app.py | |
| pinned: false | |
| hf_oauth: true | |
| short_description: replace video background with LTX alpha LoRA + H3 inpainting | |
| suggested_hardware: zero-a10g | |
| # Video Replace Background | |
| Put a video subject on a new, generated background - or just remove the background - by chaining two models: | |
| 1. **Lift the subject** with the LTX-2.5 Alpha-Gen matte LoRA ([Lightricks/LTX-2.5-22b-IC-LoRA-Alpha-Gen](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Alpha-Gen)), | |
| called over the gradio API: video in, alpha matte + green-screen out. | |
| 2. **Regenerate the background** with MiniMax-H3 ref2va inpainting ([multimodalart/MiniMax-H3-Pruned](https://huggingface.co/multimodalart/MiniMax-H3-Pruned)), | |
| held in this Space via the [inpainting modular blocks](https://huggingface.co/diffusers-modular/minimax-h3-inpainting). | |
| The subject is handed to H3 as conditioning; the inverted matte marks the background as the region to repaint. Text | |
| encoding runs on the `qwen3vl-conditioner` Space, so this one never loads a 60 GiB encoder. | |
| Upload any clip and describe the background, or turn on **Just remove the background** for a subject-on-black cutout | |
| (no generation). **Relight strength** is the quality lever: at 0 the subject is held exactly (best when the new | |
| background's light matches the original); turn it up and H3 lights the subject into the new scene so it does not read | |
| as a cutout, while keeping identity. | |
| Runs on ZeroGPU with the [ref2v turbo LoRA](https://huggingface.co/lightx2v/Minimax-h3-Turbo) at 8 steps. The matte | |
| and conditioner calls bill the caller's ZeroGPU quota, same as the denoise. The matte is cached per input clip, so | |
| changing only the prompt reuses it and skips the LTX pass. | |