Initial commit

2026-05-04 09:38:56 +02:00
commit f35e301ef7
16 changed files with 1774 additions and 0 deletions
@@ -0,0 +1,61 @@
+# Kawai
+
+Local AI image and video generator. Simple UI. NSFW-capable. Auto GPU detection (Nvidia / AMD / Intel / Apple Silicon / CPU).
+
+## Quick start
+
+```
+python launcher.py
+```
+
+Run with **any Python** you have installed. The launcher bootstraps `uv` and uses it to fetch a clean Python 3.11 runtime + venv, then installs the right PyTorch build for your GPU. Nothing about your system Python is touched.
+
+Works on Windows, Linux, and macOS.
+
+First run takes a few minutes (uv install + Python 3.11 download + torch + dependencies). Subsequent runs start instantly.
+
+### Force a specific backend
+
+Auto-detect picks one of `cuda` (NVIDIA), `rocm` (AMD on Linux), `directml` (AMD/Intel on Windows), `mps` (Apple Silicon), or `cpu`. To override:
+
+```
+python launcher.py --backend cuda     # force CUDA wheel
+python launcher.py --backend rocm     # AMD on Linux (ROCm)
+python launcher.py --backend directml # AMD/Intel on Windows
+python launcher.py --backend mps      # Apple Silicon (macOS)
+python launcher.py --backend cpu      # CPU only
+python launcher.py --reinstall        # wipe install marker, re-detect, reinstall torch
+```
+
+`--vendor {nvidia,amd,intel,cpu}` is available too if you need to pair (e.g. `--backend directml --vendor intel`). Override is persisted in `config.local.json` and survives relaunches until you pass `--backend` again or `--reinstall`.
+
+### What the launcher does
+
+1. Installs `uv` to `.tools/` if not present.
+2. Creates `venv/` with Python 3.11 (uv downloads the interpreter on demand).
+3. Detects GPU (Nvidia / AMD / Intel / Apple Silicon / CPU) and installs matching PyTorch wheel.
+4. Installs latest `diffusers`, `transformers`, etc.
+5. Opens browser UI at `http://127.0.0.1:7860`.
+
+### Reset
+
+Delete `venv/` and `.tools/` to force a clean reinstall.
+
+## Hardware tiers
+
+| VRAM | Default image model | Default video model |
+|------|--------------------|---------------------|
+| 4 GB | SDXL Turbo (fp16) | disabled |
+| 8 GB | Pony Diffusion XL | LTX-Video (fp8) |
+| 12 GB | Illustrious XL | LTX-Video (fp16) |
+| 16 GB+ | Illustrious XL + refiner | Wan 2.1 |
+
+User can override defaults in UI.
+
+## Safety
+
+CSAM detection on all outputs (NudeNet age classifier + hash check). All other content allowed: NSFW, gore, violence.
+
+## Status
+
+Windows + Linux + macOS. AMD on Linux uses ROCm; AMD/Intel on Windows use DirectML; Apple Silicon uses MPS. Intel Macs run on CPU only (no GPU acceleration path).