feat: package isolated two-step Kaggle GPU verification

fix: restore pipeline correctness and gate training on verified data

fix(scheduler): patch BasicSR CosineAnnealingRestartLR out-of-bounds crash on resume

fix(pipeline): force zero distortion on pure quality, fix radio state sync, and extend training to 6000 iters

fix(kaggle): prune intermediate checkpoints and increase save_checkpoint_freq to prevent disk full

fix(kaggle): resolve CUDA OOM, KeyError tensor(32) in ONNX export, and checkpoint export path

feat: v3 ONNX support, ArcFace paths, CPU batch fix, kaggle runner, get_available_models UI

fix: resolve GPU training bottleneck (num_gpu=1, batch_size=4, prefetch_mode=cuda, workers=4)

fix: reset resume_state to null for clean cloud GPU training and fix export command

fix: ensure basicsr/version.py is tracked and generated properly

chore: clean up temporary log files, test images, and cache

feat: enforce held-out benchmark workflow

feat: complete Phase 6 Stage II Transformer fine-tuning setup and verification with exit code 0

feat: verify Phase 7 A/B Model & Preset Pipeline with exit code 0

feat: verify Phase 3 ArcFace Identity Loss integration with exit code 0

feat: add 20 specialized project skills for image quality enhancement and CPU performance

feat: integrate Wink Quality Engine, skin grain retention, eye/lip sparkle, and update Obsidian vault

feat(model): integrate ArcFace identity loss into CodeFormer training pipeline (Phase 3)

fix(hf): integrate Static INT8 ONNX build into Dockerfile and fix thread state guard in app.py

fix: add gc.collect() in train.py to prevent memory leak during long CPU runs

fix: train_custom.py never writes to original yml — use temp file to prevent total_iter corruption

feat: add sequential model improvement roadmap (Phase 1 complete — yml config fixes)

feat: optimize training parameters, fix pipeline sequential bug, update UI default fidelity, and configure HF Spaces auto-deployment in Docker mode

Fix training crash on CPU, reduce val_freq, add OpenCV env vars

CPU optimizations and UI defaults

feat: optimize CPU performance and add agents rules

feat: add Google Colab GPU notebook, ONNX export script, and pipeline ONNX optimization

feat: extend CodeFormer training from iter 1000 to 2000

fix: resume from iter 915 and reduce checkpoint freq to 50 to prevent disk overflow

chore: bump total_iter to 1000 for CodeFormer and Real-ESRGAN

feat: add multi-category dataset crawler + Real-ESRGAN training pipeline

fix: resolve basicSR training loop termination bug when changing dataset size

merge: resolve conflicts in train_custom.py and CodeFormer_stage3_custom.yml

feat: add custom stage 3 training runner and fix Windows/CPU resume path compatibility in BasicSR

feat: implement custom training configurations and runner script for model improvement

feat: add CPU-based toy training configs and fix net_g.module AttributeError for CPU training

fix: resolve NameError in app.py and upgrade Gradio APIs in web-demos

feat: clone CodeFormer, download pretrained weights, and verify imports