Repository navigation
LoRA Pilot v2.6
LoRA Pilot v2.6 brings guided training, checkpoint comparisons, an integrated caption editor and optional MCP access into a refreshed ControlPilot workspace.
Highlights
- Added optional scoped MCP access with guided setup, dataset and run permissions, expiring credentials, and approval controls for training, comparisons and exports.
- Redesigned ControlPilot's Dashboard, Settings and Services, with grouped navigation, mobile theme controls, shared service definitions and direct tool links.
- Added build identity, sanitized support diagnostics, a collapsible global activity indicator and saved training drafts.
- Integrated Caption images into ControlPilot's light and dark themes, with dataset imports, bulk tagging and captioning, dedicated settings, 50-image pages and keyboard preview navigation.
- Added dataset previews, caption coverage, quality reports and an orange robot first-LoRA guide.
- Added guided FLUX.1 dev training alongside SDXL, a persistent training queue, searchable history, checkpoint recovery, downloads and portable experiment exports.
- Added training progress estimates, actionable failure messages and editable GPU suggestions informed by completed local runs. Training and ComfyUI can run concurrently.
- Added Move and Copy actions for the LoRA library and comparison grids covering all checkpoints; fixed prepared workflows opening empty in ComfyUI.
- Migrated RunPod shutdown to REST v2 and added optional storage allocation and spending information, with credentials kept on the backend.
- Added visual storage usage and reviewed cleanup of completed training files; corrected workspace capacity reporting on shared RunPod storage.
- Made VS Code Server an optional, persistent install from Services and improved TensorBoard run discovery and status reporting.
- Bundled pinned ComfyUI-GGUF and VideoHelperSuite extensions, including OpenCV and imageio-ffmpeg dependencies.
- Fixed ComfyUI startup imports and direct-port access protection, SDXL tokenizer caching and VAE precision, and Diffusion Pipe output-path validation.
- Standardized screen request cancellation and stale-response handling; fixed the Models page HTTP 500 error and broken documentation images.
- Updated setup, training, API and MCP documentation, with detailed validation records kept separate from release notes.
Upgrade notes
Keep your persistent /workspace volume when replacing the container. VS Code Server is now an optional install from ControlPilot → Services; existing settings and extensions remain in the workspace.
MCP is disabled by default. Configure access in Settings and grant each connection only the datasets, runs and operations it needs. See the MCP guide for setup and current integration limits.
RunPod storage and billing information require the corresponding API permissions. Missing permissions leave these features unavailable without blocking local tools.
Verification and images
Release preparation passed 408 Python tests and 61 JavaScript tests. GitHub reported no open code-scanning alerts. GPU training and generation were not rerun during release preparation.
Pushing the release tag triggers the Docker publishing workflow. Image publication is separate from this GitHub release; check the workflow result before deploying a new image. The stable Docker tag is managed separately.