Sync install section with voxtell 0.1.2 on PyPI

README.md

@@ -96,21 +96,27 @@ conda activate voxtell
 
 ### 2. Install PyTorch
 
-> [!WARNING]
-> **Temporary Compatibility Warning**  
-> There is a known issue with **PyTorch 2.9.0** causing **OOM errors during inference** (related to 3D convolutions — see the PyTorch issue [here](https://github.com/pytorch/pytorch/issues/166122)).  
-> **Until this is resolved, please use PyTorch 2.8.0 or earlier.**
-
 Install PyTorch compatible with your CUDA version. For example, for Ubuntu with a modern NVIDIA GPU:
 
 ```bash
-pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu126
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
 ```
 
+Any PyTorch >= 2.1.2 works except the 2.9.x series, which has a 3D-convolution memory
+regression (also excluded by nnU-Net).
+
 *For other configurations (macOS, CPU, different CUDA versions), please refer to the [PyTorch Get Started](https://pytorch.org/get-started/previous-versions/) page.*
 
-Install the latest version directly from the repository (you can also use
-[uv](https://docs.astral.sh/uv/)):
+### 3. Install VoxTell
+
+Install the latest release from PyPI (you can also use [uv](https://docs.astral.sh/uv/)):
+
+```bash
+pip install voxtell
+```
+
+For the development version (the `main` branch, which may contain unreleased changes), install
+directly from GitHub:
 
 ```bash
 pip install git+https://github.com/MIC-DKFZ/VoxTell.git
@@ -191,10 +197,11 @@ voxtell-predict -i case001.nii.gz -o output_folder -p "liver" "spleen" --save-co
 
 | Argument | Short | Required | Description |
 |----------|-------|----------|-------------|
-| `--input` | `-i` | Yes | Path to input NIfTI file |
+| `--input` | `-i` | Yes (unless `--jobs`) | Input NIfTI file(s), or a single folder of NIfTI files |
 | `--output` | `-o` | Yes | Path to output folder |
 | `--model` | `-m` | No | Path to a local model directory. If omitted, uses `VOXTELL_MODEL` or downloads the default model (`voxtell_v1.1`) from Hugging Face |
-| `--prompts` | `-p` | Yes | Text prompt(s) for segmentation |
+| `--prompts` | `-p` | Yes (unless `--jobs`) | Text prompt(s) for segmentation, applied to every input image |
+| `--jobs` | | No | Per-image prompts instead of `-i`/`-p`: a `.json` file or inline JSON list of `{"image": ..., "prompts": [...]}` |
 | `--device` | | No | Device to use: `cuda` (default) or `cpu` |
 | `--gpu` | | No | GPU device ID (default: 0) |
 | `--save-combined` | | No | Save multi-label file instead of individual files |
@@ -374,6 +381,28 @@ for i, prompt in enumerate(text_prompts):
 napari.run()
 ```
 
+## 🌐 Remote Inference Server
+
+Run VoxTell as an HTTP server so a lightweight client (e.g. the
+[napari-voxtell](https://github.com/MIC-DKFZ/napari-voxtell) plugin on a laptop) can drive
+inference on a remote GPU machine. The client uploads a NIfTI once; the server reorients it,
+runs sliding-window inference, streams per-patch progress, and returns the masks.
+
+```bash
+pip install "voxtell[server]"
+
+# Bind to localhost (reach it over an SSH tunnel) - safe default for a single user
+voxtell-server --host 127.0.0.1 --port 1527
+
+# ...or expose it on a trusted LAN behind a static bearer token
+voxtell-server --host 0.0.0.0 --port 1527 --api-key "$VOXTELL_API_KEY"
+```
+
+Model resolution matches `voxtell-predict` (`-m/--model` → `$VOXTELL_MODEL` → Hugging Face
+download). From a laptop, forward the port with `ssh -N -L 1527:127.0.0.1:1527 workstation`
+and point the client at `http://127.0.0.1:1527`. The client/server design is inspired by
+MIC-DKFZ's [nnInteractive](https://github.com/MIC-DKFZ/nnInteractive) (Apache-2.0).
+
 ## 🎯 Fine-Tuning
 
 Transfer VoxTell's pretrained image **encoder** into nnU-Net and fine-tune it for