faceless / README.md
Isaacdnigerian's picture
Update README.md
f961715 verified
|
Raw History Blame Contribute Delete
3.08 kB
---
title: Faceless
emoji: πŸ“ˆ
colorFrom: pink
colorTo: indigo
sdk: gradio
sdk_version: 6.20.0
python_version: '3.12'
app_file: app.py
pinned: false
license: mit
short_description: video gen pipeline
---
# Faceless LTX-Video (HF ZeroGPU Space)
Free-GPU video backend for the faceless-reels pipeline. Runs LTX-Video on
ZeroGPU and exposes a Gradio API function `generate` that returns the clip as
base64, which the pipeline decodes in `src/lib/video.ts`.
## Create the Space
1. On huggingface.co β†’ **New β†’ Space**.
- **SDK: Gradio**
- **Hardware: ZeroGPU** (this is the free on-demand GPU; you said you can
create one).
2. Add these files to the Space repo (drag-drop in the web UI, or `git push` to
the Space's git remote):
- `app.py` (from this folder)
- `requirements.txt` (from this folder)
- this `README.md` (its frontmatter configures the Space)
3. The Space builds and, on first boot, downloads the LTX weights (a few
minutes). When it's "Running", note its URL, e.g.
`https://<user>-faceless-ltx-video.hf.space`.
## Point the pipeline at it (Heroku Config Vars)
- `HF_SPACE_URL` = the Space base URL above (no trailing slash)
- `HF_TOKEN` = a token from huggingface.co β†’ Settings β†’ **Access Tokens**
(recommended β€” attributes ZeroGPU usage to your quota; required if the Space
is private)
- `VIDEO_MIN_INTERVAL_MS=0`
- `VIDEO_RESOLUTION=704x1280`
- `MAX_CLIPS=2` (ZeroGPU has a daily GPU-time quota β€” keep clip counts modest)
Also make sure the paid self-host vars are unset so this backend is selected:
`CEREBRIUM_VIDEO_URL`, `RUNPOD_ENDPOINT_ID`, `MODAL_VIDEO_URL` empty.
## Verify directly first
```bash
# 1) submit
curl -sS -X POST "$HF_SPACE_URL/gradio_api/call/generate" \
-H 'Content-Type: application/json' \
${HF_TOKEN:+-H "Authorization: Bearer $HF_TOKEN"} \
-d '{"data":["a neon city street at night, camera slowly pushing in",704,1280,5]}'
# -> {"event_id":"abc123"}
# 2) fetch the result stream (replace EVENT_ID)
curl -sN "$HF_SPACE_URL/gradio_api/call/generate/EVENT_ID" \
${HF_TOKEN:+-H "Authorization: Bearer $HF_TOKEN"}
# -> event: complete
# data: ["<base64 mp4>"]
```
## Contract (matches src/lib/video.ts)
```
POST {HF_SPACE_URL}/gradio_api/call/generate
{ "data": [prompt, width, height, duration] } -> { "event_id": "..." }
GET {HF_SPACE_URL}/gradio_api/call/generate/{event_id}
-> SSE ending in: data: ["<base64 mp4>"]
```
## Notes
- **ZeroGPU quota**: free ZeroGPU has a daily GPU-seconds limit. LTX clips are
short (tens of seconds of GPU each), so a handful of clips/day fits; big
batches may exhaust it. Keep `MAX_CLIPS` low.
- **Per-call cost**: the pipe is moved to CUDA inside `@spaces.GPU` on each call
(ZeroGPU only exposes the GPU there). That adds a few seconds; fine for short
clips. If you hit the 120s `duration` cap, lower resolution or `num_frames`.
- **If the build errors on `pkg_resources`** (as Cerebrium did), add
`setuptools<80` to `requirements.txt`.
- **Model revision**: pin `MODEL_ID` in `app.py` for reproducible weights.