docs: add CUDA tag to README, getting-started, deployment, benchmarks

- README: add CUDA docker run example alongside full and lite
- Getting started: add GPU acceleration tip with speedup numbers
- Deployment: add CUDA row to variants table
- Docker tags: expand benchmarks with warm + cold start tables

apps/docs/guide/deployment.md

@@ -2,12 +2,13 @@
 
 Stirling Image ships as a single Docker container. The image supports **linux/amd64** and **linux/arm64**, so it runs natively on Intel/AMD servers, Apple Silicon Macs, and ARM devices like the Raspberry Pi 4/5.
 
-Two variants are available:
+Three variants are available:
 
 | Variant | Tag | Size | What's included |
 |---------|-----|------|-----------------|
 | Full | `:latest` | ~11 GB | All tools + AI/ML (background removal, upscaling, OCR, face blur, object eraser) |
 | Lite | `:lite` | ~1.5 GB | All image processing tools, no AI/ML |
+| CUDA | `:cuda` | ~14 GB | Full + GPU-accelerated AI (NVIDIA only, amd64) |
 
 See [Docker Tags](./docker-tags) for the full comparison, Docker Compose examples, and version pinning.
 

apps/docs/guide/docker-tags.md

@@ -45,18 +45,32 @@ Same tools as the full image, but built with GPU-accelerated Python packages (on
 
 Requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) on the host. Linux amd64 only.
 
-### What gets faster with GPU
+### Benchmarks
 
-| Tool | CPU | GPU (RTX 4070) | Speedup |
-|------|-----|----------------|---------|
-| Background removal | 2.4s | 0.9s | 2.7x |
-| Upscale 2x | 350ms | 310ms | ~1.1x |
-| Upscale 4x | 910ms | 310ms | ~3x |
-| OCR (PaddleOCR) | 137ms | 94ms | ~1.5x |
+Tested on an NVIDIA RTX 4070 (12 GB VRAM) with a 572x1024 JPEG portrait. Both images ran on the same machine. "Warm" means the model is already loaded in memory (second request onward).
 
-Benchmarked with a 572x1024 JPEG portrait. Larger images show bigger speedups, especially for upscaling.
+#### Warm performance
 
-Non-AI tools (resize, crop, convert, etc.) are unaffected since they use Sharp (CPU-based).
+| Tool | CPU | GPU | Speedup |
+|------|-----|-----|---------|
+| Background removal (u2net) | 2,415ms | 879ms | 2.7x |
+| Background removal (isnet) | 2,457ms | 1,137ms | 2.2x |
+| Upscale 2x | 350ms | 309ms | 1.1x |
+| Upscale 4x | 910ms | 310ms | 2.9x |
+| OCR (PaddleOCR) | 137ms | 94ms | 1.5x |
+| Face blur | 139ms | 122ms | 1.1x |
+
+#### Cold start (first request after container start)
+
+| Tool | CPU | GPU | Speedup |
+|------|-----|-----|---------|
+| Background removal | 22,286ms | 4,792ms | 4.7x |
+| Upscale 2x | 3,957ms | 2,318ms | 1.7x |
+| OCR (PaddleOCR) | 1,469ms | 1,090ms | 1.3x |
+
+Cold start includes loading the model into memory. GPU cold starts are faster because CUDA parallelizes the model loading.
+
+Larger images show bigger speedups, especially for upscaling. Non-AI tools (resize, crop, convert, etc.) are unaffected since they use Sharp (CPU-based).
 
 ### GPU health check
 

apps/docs/guide/getting-started.md

@@ -24,6 +24,16 @@ stirlingimage/stirling-image:lite
 All 27+ image processing tools work the same. See [Docker Tags](./docker-tags) for the full comparison.
 :::
 
+::: tip GPU acceleration
+Have an NVIDIA GPU? The CUDA image auto-detects your GPU and accelerates background removal (2.7x), upscaling (3x), and OCR (1.5x):
+
+```bash
+docker run -d --gpus all -p 1349:1349 -v stirling-data:/data stirlingimage/stirling-image:cuda
+```
+
+Requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html). Falls back to CPU if no GPU is found. See [Docker Tags](./docker-tags) for details and benchmarks.
+:::
+
 ## Run with Docker Compose
 
 Create a `docker-compose.yml`: