Model Management
Install the selected model, understand readiness and recover downloads safely
Last updated: 2026-10-27
The model catalogue separates what is listed, what is installed, what fits your hardware and what is currently loaded. A catalogue entry or a completed transfer alone is not proof a render can start.
Official Model Catalog
Use the app's Models / Download Models catalogue for current entries, sizes, release labels and install requirements. Pipelines distinguishes production diffusion families, Alpha tools and non-diffusion utilities. Counts and download sizes change; this guide does not keep a second model inventory.
Besides base models, a task can require LoRA, ControlNet, IP-Adapter, vision, matte, depth or upscaling weights. Install the selected task's displayed requirements rather than every optional companion. 360° Depth is a utility for reusable depth maps, not a diffusion pipeline; see Channels.
Installing Models
First-Run
The setup wizard offers model installation. Review the selected members, drive space and hardware guidance. A recommended group can partly succeed: inspect each item if the overall request reports an issue.
After First Run
- Open the model catalogue from Download Models or the selected pipeline/source's requirements action.
- Choose the exact pipeline and model, including supported precision/speed options where exposed. A similarly named model from another family is not interchangeable.
- Read the GPU fit and required companion files. Fit is an estimate, not a reservation of currently free memory or a performance guarantee. Unknown fit needs investigation, not an assumption that it fits.
- Click Install and inspect Downloads. A queued acknowledgement means the request was received, not that files are ready.
- Wait for download and verification to settle. Return to the source/model readiness display before generating.
Download Queue
The queue has one active transfer slot; pending work can wait behind it. Priority moves pending work ahead, without starting an independent transfer. Progress can show measured bytes/rate or an indeterminate phase when the total is unknown. It must not be read as a guaranteed completion time.
Cancel requests a stop. Wait for the active writer to settle before another install, repair, uninstall or storage move. Matching partial files can be reused by managed resume; cancelled/failed work does not imply existing model files were deleted. After restart, inspect recovered entries and any attention message. Do not promise every interrupted request resumes automatically.
Gated Models (HuggingFace Authentication)
Some models require a model-host account, accepted access terms and a token. Follow the app's authentication/access dialog, accept the exact repository's terms with the same account, then retry the selected request. An authentication error differs from missing space, unavailable source or incompatible hardware. Never put access tokens in public logs or support screenshots.
Managing Installed Models
Use the installed model's own actions. Uninstall removes installed files while keeping an official catalogue entry available for later installation. Custom registration removal is distinct and can have file effects; read its confirmation and outcome. Do not manually erase a shared LoRA/model folder to remove one item.
For a failure, inspect readiness and the offered Repair action. A load error alone does not prove corrupt weights: access, runtime, storage, provenance and source conflicts can need different recovery. Repair verifies selected replacement files; an incomplete repair can keep the model blocked until inspection is healthy. Do not bypass the app's held/incomplete state by editing files.
VRAM Note
Disk installation and GPU residency are different. Selecting an installed model may still require loading/unloading and can fail if another task owns the GPU. CPU offload uses system RAM and takes time; precision options are model-specific. See GPU Management, System Requirements and Parameters.
Backup and offline use
Keep enough free space for the selected artifacts, temporary/staged work and outputs; a catalogue size is an estimate. Storage location follows your app data-root settings. A saved Workflow does not contain weights or access credentials. Installed local models can be used without downloading again, but license checks and optional online services retain their own requirements; see License Activation.