LoRA Pilot v2.5.8
LoRA Pilot v2.5.8 brings the Models workflow catalog, reviewed video-model installation, optional ComfyUI access protection, refreshed bundled applications, and startup and download reliability fixes.
These notes compare against the previous GitHub release, v2.5.4. They include the v2.5.5–v2.5.7 changes recorded in the changelog, which did not have separate GitHub releases.
Release Scope
- Full release tag:
v2.5.8 - Baseline tag:
v2.5.4 - Commits covered since
v2.5.4: 40 - Open GitHub code-scanning alerts at release preparation: 0
- Windows preview releases remain a separate release line.
Highlights
Models and Video Workflows
- Redesigned Models around a searchable family/workflow catalog with task filters, a details panel, Installed models, and Downloads with progress and retries.
- Added reviewed installation for the bundled LTX-2.5 and MiniMax H3 text-to-video and image-to-video workflows. Review exact required files, destinations, source access, download sizes, and free disk space before queuing missing components.
- Reuse completed files and active download jobs; the LTX prompt enhancer remains optional.
- Bundle four official workflow graphs with provenance and licensing, and preserve existing workspace workflow edits. Model weights download separately.
- Repaired model sources, names, sizes, and canonical destinations. Downloads no longer repeat upstream directory segments, and conflicting filenames use model-specific locations.
- Added repository completion records and file-size checks so incomplete downloads are not presented as installed. Verified legacy files can be reused without discarding user data.
- Extracted Models routes and download queue handling from
app.pyinto dedicated modules while preserving the existing API behavior.
ControlPilot, MediaPilot, and Training Reliability
- Added optional ComfyUI UI/API protection through ControlPilot sessions and revocable API tokens, including HTTP and WebSocket forwarding. Protection is off by default.
- Extended ControlPilot protection to Diffusion Pipe and ComfyUI proxy routes, and hardened dataset, model, and training path handling.
- Fixed MediaPilot move collisions, preserved TrainPilot move results, and corrected TagPilot text/caption decoding.
- Fixed Diffusion Pipe startup races and log handling, Copilot timeout handling, shutdown concurrency, and runtime lifecycle failures.
- Disable Copilot input while its sidecar is unavailable and keep availability synchronized with service actions.
- Improved Models mobile overflow and MediaPilot embedding on desktop and mobile.
Bundled Apps and CUDA Environments
- Updated ComfyUI to
v0.34.0, Kohya SS tov26.0.0, InvokeAI to6.14.0, and AI Toolkit to commitb36bb3998ae596a566d85513299696a3a78f0dcb. - Updated code-server to
4.135.0, JupyterLab to4.6.3, and refreshed supporting runtime dependencies. - Keep CUDA 13.0 as the default and CUDA 12.8 as an alternative profile. Core PyTorch remains
2.11.0. - Isolated Kohya, InvokeAI, and AI Toolkit dependencies from core. TrainPilot uses the Kohya environment; InvokeAI retains its separate PyTorch
2.7.1+cu128stack on both profiles. - Added system FFmpeg/shared libraries, AI Toolkit SQLite build compatibility, and expanded dependency and GPU smoke checks.
Startup Persistence and Security
- Preserve custom Supervisor configuration and optional workspace ComfyUI environments across startup. Missing custom Supervisor configurations fail explicitly.
- Refresh bundled apps and documentation with an inventory that preserves user modifications and runtime data, without creating legacy backup copies.
- Serialize bootstrap credentials safely, including multiline values, and preserve unrelated settings across repeated starts.
- Connect Supervisor authentication to the exported generated or supplied password.
- Keep rewritten MediaPilot configuration private with
0600permissions and secure temporary files. - Expanded regression coverage across startup, access protection, model installation, APIs, downloads, and runtime lifecycle behavior.
Documentation and Publishing
- Updated model management, ComfyUI access, CUDA compatibility, component, and RunPod/CLI documentation.
- Added contribution and review guidance and a Docker Hub publishing workflow for version tags.
Upgrade Notes
- Build the image from
v2.5.8. The tag's Docker publishing workflow failed at startup before creating any jobs, so no versioned image was published by this run. Docker workflow status. - Keep the persistent
/workspacevolume when replacing the container. Customized files and completed model downloads remain workspace-backed. - Existing repository downloads without completion records need another
models pull <name>to verify cached files and record completion. Corrected model destinations may also require a new pull; legacy files are preserved. - Download queues are process-local. After a Portal restart, review the workflow again to queue remaining files.
- ComfyUI protection is opt-in. Configure a ControlPilot password before enabling it in Settings.
- If you maintain a custom Supervisor configuration with the old shell-style password placeholder, update it to
%(ENV_SUPERVISOR_ADMIN_PASSWORD)s; custom configuration files are preserved. - Use
CUDA_PROFILE=cu128when building the alternative CUDA 12.8 profile.
Known Notes
- No prebuilt Docker image or Windows installer is attached to this GitHub release.
- Workflow/model installation checks do not prove successful GPU generation. Third-party custom nodes remain user-managed.
- Full image builds and live GPU/RunPod validation were not performed during this release preparation.
Verification
- Python regression suite:
.venv/bin/python -m unittest discover -s tests— 166 tests passed. bash -n scripts/bootstrap.sh— passed.- Supervisor's actual configuration parser verified generated and supplied passwords.
git diff --check— passed.- GitHub reported 0 open code-scanning alerts at release preparation; this is not a claim of a completed scan of the new release commit.
make build-checkcould not run because the local Docker daemon socket was unavailable.
Compare
- Changes since the previous GitHub release: v2.5.4...v2.5.8