Set up local models

Python packages (the Requirements dialog)

The image-gen stack (PyTorch, Diffusers, Transformers) is not bundled in the installer; it’s several gigabytes and only image-gen needs it. The first time you use image generation, the in-app Requirements dialog installs the packages into a private runtime at %LOCALAPPDATA%\ReelMarkr\python_runtime\. You can remove the AI stack without touching your ReelMarkr install.

Removing packages

Remove packages in the Requirements dialog, not by deleting the runtime folder by hand:

• Uninstall a single package from its own row.
• Remove all clears every optional package in one action.
• Right-click the Runtime path to open the packages/runtime folder on disk.

A row flips back to Install as soon as the package is removed, even if it’s still loaded in the running session.

Use it when: one bad Torch build needs a reinstall, not the whole runtime.

Image-generation models

Download models from Civitai or HuggingFace inside the app. Models are plain files on disk, and you can point ReelMarkr at a custom models folder in the image-gen settings.

AI assistant backend

The AI assistant talks to a local LLM server. Install one of these once. Both are free and need no account:

Ollama

1. Download from ollama.com/download and install.
2. Ollama runs in the background at 127.0.0.1:11434.
3. In Chat settings, choose Ollama and pick a model, or open the Models tab and download one in-app (Qwen 2.5, Llama 3.1/3.2, Gemma 3, and the LLaVA vision models, with sizes). You can also pull models yourself: ollama pull qwen2.5:7b-instruct.

LM Studio

1. Download from lmstudio.ai and install.
2. In LM Studio, download a model, then start its Local Server (defaults to 127.0.0.1:1234).
3. In Chat settings, choose LM Studio and select the loaded model.

ReelMarkr auto-detects both on the default ports. For a server on another machine, set the Host/Port in Chat settings.

Dictation & voice-memo model

Dictation and voice-memo transcription use a local Whisper (faster-whisper) model on your CPU. No GPU, and no audio leaves your machine. The package (~150 MB) and the model download automatically on first use. Sizes range from Tiny (75 MB) to Large (2.9 GB); Base (140 MB) is the default. Manage or remove them under the Immersion Canvas Settings ▸ Extras tab.

Settings ▸ Speech processing device runs Whisper on the CPU (default) or a specific CUDA GPU, which speeds up long voice-memo transcriptions.

GPU notes (image generation)

Image generation runs on your GPU: an NVIDIA card with recent drivers on Windows, Metal (MPS) on Apple Silicon Macs. Speed and maximum model size scale with VRAM:

• 6-8 GB: SD 1.5 comfortably; SDXL is possible but slower.
• 10-12 GB: SDXL at reasonable speed, ControlNet, upscaling.
• 16 GB+: SDXL with heavier stacks (multiple ControlNets, video models).

Less VRAM still works, but slower, and it caps resolution and model size. Dictation and the assistant don’t need a GPU.