Model Assets

This page covers how Local Dream handles checkpoints, LoRAs, and related assets.

Custom Models

You can use your own models in addition to the built-in ones:

• CPU/GPU mode: Import any SD1.5 checkpoint directly in the app. The conversion happens on-device. CPU models can also use OpenCL acceleration where supported.

  WARNING

  The checkpoint must include VAE weights (typically ~2 GB or larger). Models without VAE weights (~1.8 GB) cannot be converted.

• NPU mode: SD1.5 and SDXL models must be converted on a host machine first. See the Conversion Guide.

For pre-converted community NPU models and details on chip tier suffixes, see Available Models — Pre-converted Community Models.

Managing Models (Pin / Rename / Delete)

Long-press a model card on the home screen to enter selection mode. The toolbar then offers, left to right:

• Pin — works on a multi-selection and toggles to Unpin when everything selected is already pinned. Pinned models sort to the top within each tab (most-recently-pinned first) and show a pin indicator, so your go-to models stay reachable as the catalog grows.
• Rename — offered only for a single selected custom model. Because a custom model's name is its on-disk id, renaming migrates everything keyed by that id together: the model directory, its history (images and records), all of its saved per-model parameters, and its pinned state. The rename is committed atomically at the directory move, so a model can never end up half-renamed.
• Delete — removes the selected models. By default a deleted model keeps its history (opt out with the checkbox), so you can reinstall later and still have your old images in the global History screen.

LoRA Support

LoRA weights cannot be attached to an already converted model, because converted Local Dream models are already quantized, both on the CPU/GPU path and the NPU path.

If you want to use a LoRA, merge it into the original checkpoint before conversion. After that, convert or import the merged checkpoint as a normal model.

In other words:

• Already converted / quantized model: no LoRA injection
• Original checkpoint before conversion: LoRA can be baked in first