studio-video / README.md
Beexly's picture
Upload folder using huggingface_hub
348bd51 verified
|
Raw History Blame Contribute Delete
2.54 kB
---
title: Beexly Studio Video
emoji: 🎬
colorFrom: red
colorTo: purple
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: other
license_name: ltx-video-open-weights
license_link: https://huggingface.co/Lightricks/LTX-Video-0.9.7-distilled/blob/main/LTX-Video-Open-Weights-License-0.X.txt
short_description: Text-to-video generation on free ZeroGPU
tags:
- zero-gpu
---
# Beexly Studio Video (ZeroGPU)
Personal video-generation tool. **Not connected to any Galaxy Sports Edge
product, data, or prediction engine β€” this is Garrett's own tooling,
unrelated to GSE.**
## Model and license
- **Model**: `Lightricks/LTX-Video-0.9.7-distilled`, loaded via diffusers'
`LTXConditionPipeline` (bf16, deferred-cuda ZeroGPU pattern).
- **License**: LTX-Video Open Weights License β€” free for personal and
non-commercial use; commercial use by an entity with $10M+ annual revenue
requires a separate paid agreement. Same posture as FLUX.1-dev on
`studio-media`: personal tool, **never for anything shipped on
galaxysportsedge.com.**
- **Why this checkpoint**: real research before building β€” the base
13B `Lightricks/LTX-Video` needs a heavier, separately-packaged
multi-scale pipeline to hit ZeroGPU's duration cap reliably (confirmed
by inspecting a working reference Space's source). The **distilled**
0.9.7 checkpoint is diffusers-native (`LTXConditionPipeline.from_pretrained`,
no custom package) and its own model card recommends just
**7 inference steps** β€” built for exactly this speed/VRAM envelope.
~10GB VRAM for the base LTX-Video pipeline per the diffusers docs.
## Known limits
- ZeroGPU's real per-call duration ceiling is ~120s reliably (up to ~240s
has been reported but isn't guaranteed) β€” this is the binding constraint
for video generation, tighter than VRAM. `num_frames` is capped at 121
(8n+1, ~5s at 24fps) and steps at 12 to stay inside it; **the
`_video_duration()` estimate is unmeasured** ahead of a real deployed
call β€” recalibrate against real wall-clock time once tested, the same
discipline used for studio-media's FLUX.1-dev steps slider.
- Same forced `gradio[oauth]==5.9.1` / `gradio_client` 1.5.2 schema bug as
every other ZeroGPU Space in this suite β€” fixed here with the same
monkeypatch (see `studio-media`'s README for the full root-cause writeup).
- Cold start re-downloads all weights (no persistent disk on free tier).
- `num_frames` must be `8n+1` (41, 49, 57, ... 121) β€” the UI slider is
stepped to only offer valid values.