A minimal, polished local UI for generating text-to-video with SkyReels-V2 (Diffusion Forcing). Runs fully on your machine (Windows/Linux), supports local model folders, and exports directly to MP4 (GPU-accelerated NVENC by default, with a safe fallback).
(https://github.com/Mosai-Sys/LAV.git)
-
Models:
DF-1.3B-540P(fast)DF-14B-540P(quality)DF-14B-720P(quality, higher VRAM)
-
Clean, dark UI inspired by skyreels.ai; all controls themed and keyboard-friendly.
-
One-click presets (“Fast 1.3B” and “Quality 14B”).
-
Advanced controls:
ar_step(Bpc),causal_block_size, overlap smoothing, add-noise condition, seed, steps, frames. -
Local model directory support (
SKYREELS_MODELS_DIR) or per-run override path. -
GPU-accelerated export via NVENC (
h264_nvenc/hevc_nvenc), with automatic fallback toimageio. -
History panel (recent videos) and Open outputs folder button.
-
Robustness: CUDA checks, OOM guidance, TF32 tuned for modern NVIDIA GPUs, consistent metadata.
- Python 3.11/3.12 (64-bit)
- NVIDIA GPU (Ampere or newer recommended). Install the latest NVIDIA driver.
- Git (to install Diffusers from source).
- Internet access (first time) to download models from Hugging Face unless you point to local copies.
RTX 50xx (Blackwell): Install a recent CUDA build of PyTorch (CUDA 12.8+) that supports sm_120. See PyTorch install below.
git clone [https://github.com/your-org/skyreels-ui.git](https://github.com/Mosai-Sys/LAV.git)
cd skyreels-ui
python -m venv venv
# Windows:
venv\Scripts\activate
# Linux/macOS:
source venv/bin/activate-
Standard CUDA GPUs (Ampere/Ada):
pip install --index-url https://download.pytorch.org/whl/cu121 torch torchvision torchaudio
-
RTX 50-series (sm_120 / CUDA 12.8+): use a recent/nightly build:
pip install --pre --index-url https://download.pytorch.org/whl/nightly/cu128 torch torchvision torchaudio
Verify:
python -c "import torch; print(torch.__version__, torch.version.cuda); print('CUDA:', torch.cuda.is_available())"
pip install -U \
git+https://github.com/huggingface/diffusers.git \
gradio accelerate transformers huggingface_hub \
imageio imageio-ffmpeg pillow safetensors ftfy \
opencv-pythonIf you want to keep models off the HF cache, place them under a folder and point the UI to it:
# Windows
setx SKYREELS_MODELS_DIR "C:\LAP\skyreels_ui\modeller"
# Linux/macOS (bash)
export SKYREELS_MODELS_DIR="$HOME/skyreels_ui/models"The app auto-detects known subfolders if they match model names/repo names.
huggingface-cli login# from the repo root
venv\Scripts\activate # or source venv/bin/activate
python skyreels_ui.pyOpen the URL shown (default http://127.0.0.1:7860).
Change port if needed:
set GRADIO_SERVER_PORT=7861 & python skyreels_ui.py # Windows export GRADIO_SERVER_PORT=7861 && python skyreels_ui.py # Linux/macOS
-
Model / Resolution: choose one of the supported SkyReels-V2 checkpoints and output resolution.
-
Prompt: describe your scene (camera motion, time of day, style/lighting).
-
Seed:
-1= random. -
Num inference steps: 24–34 is a good range.
-
Frames (length): more frames = longer clip.
-
Overlap history: smooths transitions between frame blocks.
-
Addnoise condition: trade-off for stability vs. detail.
-
Advanced (Bpc):
ar_step> 0 enables async/Bpc mode.- If
causal_block_sizeis 0, it defaults toar_step.
-
Override model: paste a HF repo id or a local path to a downloaded checkpoint folder.
Output: MP4 is saved to outputs/skyreels_<model>_<res>_<timestamp>.mp4.
Use “Open outputs folder” or the History dropdown to preview recent renders.
-
Quick test (1.3B @ 540p)
steps=26–30,frames=197–257,ar_step=0,overlap=17,addnoise=20. -
Longer & stable (1.3B @ 540p)
steps=30–34,frames=377–513,ar_step=2,causal_block_size=2,overlap=17,addnoise=20.
Performance notes
- TF32 is enabled for conv + matmul via the new PyTorch API.
- Optional (memory):
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True.
The app first tries NVENC (h264_nvenc / hevc_nvenc) for fast GPU video encoding, then falls back to imageio if unavailable.
Ensure these are installed in the same venv:
pip install -U imageio imageio-ffmpeg opencv-pythonYou should see Video Encode activity in Task Manager (Windows) during export.
-
Port already in use (7860): set
GRADIO_SERVER_PORT(see Run). -
HF 401/403: run
huggingface-cli login. -
hf_transferwarning: the app turns it off by default. To override:- disable:
set HF_HUB_ENABLE_HF_TRANSFER=0
- disable:
-
Windows symlink warning: harmless. Enable Developer Mode or run Python as admin to suppress it.
-
CUDA OOM: reduce
frames,steps, setar_step=0, or use lower resolution. -
“Diffusers missing SkyReelsV2 classes”: ensure you installed diffusers from GitHub (see step 3).
-
No MP4 written / export error: Install
imageio imageio-ffmpeg(andopencv-pythonfor NVENC). The UI will display a clear error if export fails.
skyreels-ui/
├─ skyreels_ui.py # the app
├─ outputs/ # rendered videos (auto-created)
├─ modeller/ # (optional) local model folders (if you use SKYREELS_MODELS_DIR)
├─ run_skyreels_ui.bat # optional launcher (Windows)
└─ README.md # this file
- Inference runs locally on your GPU.
- Hugging Face token (if used) is only for downloading gated/large models and is stored by
huggingface_hub.
- SkyworkAI – SkyReels-V2 (Diffusion Forcing) models
- Hugging Face – Diffusers, Transformers, Hub
- PyTorch — CUDA/TF32 goodness
- Gradio — modern, fast UI toolkit
- imageio / imageio-ffmpeg / OpenCV — reliable video export
Choose a license for this repository (e.g., MIT). Example:
MIT License
Copyright (c) 2025 ...
Issues and PRs are welcome.
If you add new models, update the MODEL_CHOICES map and consider adding a preset (height/width/frames).
Add a screenshot (e.g., docs/screenshot.png) showing the UI while rendering.
Q: Can I run fully offline?
Yes. Download the model folders once (or on another machine), place them under SKYREELS_MODELS_DIR, and use the local path.
Q: Does it support Linux? Yes—tested with recent NVIDIA drivers and CUDA builds of PyTorch.
Q: Why no audio?
This app focuses on visual generation/export. You can mux audio later (e.g., ffmpeg -i video.mp4 -i audio.mp3 -shortest out.mp4).
Happy rendering!