Known Issues

A short list of quirks that come up often. If you hit them, the workaround is usually a setting on your device rather than a fix in the app.

Local HTTP Communication Can Be Blocked

Local Dream's UI talks to its on-device C++ model backend over HTTP on 127.0.0.1:8081 (the same backend documented in HTTP API Reference). Anything that interferes with localhost HTTP traffic will prevent generation from starting.

Common culprits:

• some VPN or proxy apps that intercept or reroute local traffic
• vendor-level "restricted network access", "frozen background apps", or aggressive battery / data policies that block the app's local sockets (common on certain OEM ROMs)

If you cannot run any model at all, check these first — disable VPNs and proxies, grant Local Dream unrestricted network access in your system settings, and try again.

MediaTek GPU: Abnormal Output

On some MediaTek devices, generating on the GPU produces garbled or otherwise broken images. This is because the OpenCL library that the GPU runtime expects cannot be found on those devices.

If you see this, switch to CPU (or NPU on supported Snapdragon hardware) for now.

Switching to Background May Stall NPU Runs

Local Dream uses an Android foreground service to run the model, so generation continues when the app moves to the background — long generations and batch jobs are designed to keep running.

However, on the NPU path some devices will stall a run if the app is switched to the background while a step is in flight. Keep Local Dream in the foreground while generating on NPU to avoid this.