Csmart Studio Desktop
Description of Release Features
Release 1.0.0-Beta.10
Subminor Version - Released on Mar 09, 2026
New Features
BioCLIP 2 backbone — Add BioCLIP 2 as a new foundation model option in the training screen
Training mode controls — Redesign weight source and training mode selection with clearer options for fine-tuning, transfer learning, and training from scratch
Find Similar screen — Search for visually similar images across the dataset to help identify duplicates or mislabeled samples
Feature Extraction screen — Extract and visualize cluster metrics from trained models to evaluate class separability
Project rename — Rename projects directly from the app, with automatic path rewriting across all references
Training profiles — Pre-configured training profiles for quick setup of common training scenarios
GPU auto-detection — Automatically detect GPU model and VRAM on startup for smarter hardware recommendations
DINOv2 Giant and EVA-02 Enormous backbones — Add larger backbone variants for higher accuracy
Export naming convention — Redesigned export naming for clearer, more descriptive model file names
Feature extraction bundle — Exported models now include the feature extraction data (NPZ) for deployment
Unsaved changes guard — Workflow screens now prompt before navigating away when changes haven't been saved
Open-source license viewer — View GPL/LGPL license texts from the License screen
Improvements
Streamlined model list — Curated backbone selection with tier metadata for easier model choice
Export report and export step order swapped for a more logical workflow
Augmentation presets reordered to match typical usage patterns
File picker defaults to the project root instead of the Desktop
Sidebar logo no longer gets cropped at smaller widths
Dataset Tools submenu restored as a collapsible group in the sidebar
Bug Fixes
Fix progress bars not updating correctly for certain long-running Python processes
Fix dataset artifacts not being cleaned up when a process is cancelled
Fix VRAM estimates for EVA-02 Enormous backbone
Fix precision setting being incorrectly reset when changing models
Fix training stderr buffer growing without limit during long sessions
Fix background throttling causing stalled processes when the app window is not focused
Release 1.0.0-Beta.9
Subminor Version - Released on Mar 03, 2026
New Features
DINOv2 and EVA-02 backbones — Add DINOv2 Base/Large and EVA-02 Base/Large as new foundation model options in the training screen
GPU memory guidance — Heavy backbones now display a warning with recommended batch sizes to help avoid out-of-memory errors
Stochastic Weight Averaging (SWA) — New callback section in Model Settings to configure SWA learning rate, epoch start, annealing epochs, and annealing strategy (cosine or linear)
Reference model support in export — Select an existing production .onnx model as a reference; on export, class mappings and configuration fields (hue colormap, defect subsets, classification methods, etc.) are automatically merged into the new model descriptor
Checkpoint summary panel — The export screen now shows a full checkpoint summary including training session results (best accuracy/loss), test session results with a session selector dropdown, and data augmentation details
Test session selector for reports — PDF report generation can now target a specific test session instead of always using the latest
Auto-derive test dataset path — The test screen automatically selects the Test subfolder from the split dataset, removing a manual configuration step
Improvements
Streamlined model selection — Removed legacy and redundant backbones (ResNet-18/34/50, ResNeXt, Wide ResNet, ConvNeXt Base, ViT Base, SegFormer B0–B4) to present a focused, curated list of recommended architectures
Mixup/CutMix config relocated — Mixup settings now correctly read from and save to data_augmentation instead of hyperparameters, matching the updated config schema
Training metrics enriched — Session metadata (session ID, start time, elapsed seconds, final status) is now saved alongside epoch-level metrics for better traceability
Export and report buttons show loading state — Visual feedback while export or report generation is in progress
Test screen simplified — Test dataset input is now read-only with a clear helper note; removed unnecessary config writes on folder selection
Release 1.0.0-Beta.8
Subminor Version - Released on Feb 28, 2026
New Features
Add a dedicated Outlier Detection screen for identifying and removing anomalous samples from datasets - Group Pre-Classification, Outlier Detection, and Cluster Dataset under a collapsible "Clean Dataset" submenu in the sidebar
Display the training hyperparameters panel on the Export screen
Track elapsed training time and number of models exported
Add unsaved changes guard to settings screens, prompting before navigating away
Auto-merge new configuration fields into existing projects when opening them
Add coffee metadata and export comments to project data
Improvements
Improve contamination parameter layout on the Outlier Detection screen
Bug Fixes
Fix "Project repaired" toast appearing every time a project is opened, even when only new default settings were added
Fix settings buttons remaining clickable when no project is open
Fix checkpoint scan failing on certain version directory names during export
Release 1.0.0-Beta.7
Subminor Version - Released on Feb 27, 2026
New Features
Add coffee metadata fields (species, variety, origin, region, processing) to projects, editable at creation time and from a new Edit Project screen
Add the issued date and user comments fields to the Export screen for documenting model exports
Generate standalone PDF reports from the Export screen with improved layout and automatic opening after generation - Add reload button for the class distribution chart on the Split Dataset screen - Add preprocess configuration section to Model Settings - Display runtime version in the Environment status message - Recover missing projects from disk via the Browse Projects screen Improvements
Cancel buttons on the Export screen are now scoped to their respective operation (Export Model vs Generate Report)
Select buttons on the Export screen are disabled while an operation is running
Bug Fixes
Fix tray icon not loading in the packaged application
Fix checkpoint version directory matching to support non-standard suffixes
Release 1.0.0-Beta.6
Subminor Version - Released on Feb 26, 2026
Generate standalone PDF reports from export results
Display a warning in the sidebar when the AI runtime is outdated.
Redesign the navigation bar with updated layout and styling
Automatically open PDF reports after generation - Improve PDF report page layout and formatting Bug Fixes
Fix runtime installations failing when interrupted by using atomic staging directory
Release 1.0.0-Beta.5
Subminor Version - Released on Feb 25, 2026
Project Management
Per-project config.yaml — each project gets its own config instead of sharing a global file; paths are stamped on creation - Browse Projects redesign — full-width cards with avatars, clickable rows, relative timestamps, and project relocation when folder moves - Project deletion lock — sidebar and controls are disabled during deletion to prevent accidental navigation - Sidebar gated by project state — workflow screens are disabled until a project is opened - Navigate to Split Dataset after project creation — starts the workflow immediately
Stability & OneDrive Resilience
OneDrive-safe filesystem operations — centralized retry engine with exponential backoff for all Python scripts; handles file locks from cloud sync, antivirus, and Explorer
Block OS shutdown during processes — prevents Windows Update restarts while training/testing/etc. is running; minimizes to tray instead of quitting
System tray integration — closing the window hides to tray when a process is active; quit via tray menu
Validate persisted paths on screen load — pre-classification, test, and export screens clear stale paths that no longer exist on disk
UI / UX
Virtualized log monitor — uses react-window for O(1) scroll performance during long runs; long lines wrap instead of truncating
Path tooltips — hovering path inputs reveals the full file path
Settings visual hierarchy — fields wrapped in teal-accented cards with helper text
Fixes
Hydra config isolation — pass --config-path to Python scripts so the template config.yaml is never overwritten
Per-project checkpoint dialogs — file pickers resolve from the active project, not stale global paths
Last updated