Add train_controlnet_image-from-mask skill

[Can-Zhao]

Skill for training the 3D image-from-mask ControlNet (scripts.train_controlnet), covering the training-data format (paired VAE latent embedding + combined integer label mask, JSON data list schema, folds, modality), the label↔latent 4x spatial relationship with nearest-neighbor resampling, configs, single/multi-GPU launch, key knobs, and outputs.

[greptile-apps]

Greptile Summary

This PR adds the train_controlnet_image-from-mask skill (new file) and extends the data-prep skill with the missing Step 4 (resample combined label to 4× the latent grid), along with a bug fix in scripts/utils.py that adds "image" to the Orientationd transform so both the latent embedding and the label are reoriented to RAS — previously only the label was reoriented, which could produce axis-misaligned pairs for non-RAS NIfTI inputs.

• scripts/utils.py: One-line fix adds "image" to Orientationd(keys=["image", "label"]), ensuring the 4-channel latent NIfTI and its paired label are always in the same canonical axis order before training.
• skills/train_controlnet_image-from-mask.md (new): Documents the full training loop — data format, JSON schema, configs, single/multi-GPU launch, version selection table, and a gotchas checklist with real shipped config defaults.
• skills/finetune_image-from-mask_data-prep.md: Adds Step 4 (nearest-neighbor resampling of the combined label to the encoded-image grid), updates step references, and cross-links to the new train skill.

Confidence Score: 5/5

Safe to merge — the utils.py fix is a correct hardening of the dataloader orientation step, and the new skill documentation accurately reflects the code.

The only executable code change adds image to an existing Orientationd transform, which is correct: MONAI handles multi-channel 4D volumes by reordering spatial axes while leaving the channel axis intact. All other changes are documentation with two minor inconsistencies that do not affect training correctness.

skills/finetune_image-from-mask_data-prep.md — Step 4 is missing an explicit save-to-disk instruction matching the pattern from Step 3.

Important Files Changed

Filename	Overview
scripts/utils.py	Bug fix: adds image to Orientationd so the 4-channel latent embedding is reoriented to RAS alongside the label.
skills/finetune_image-from-mask_data-prep.md	Adds Step 4 for resampling the combined label; missing save instruction and stale Step 3 comment in JSON schema.
skills/train_controlnet_image-from-mask.md	New skill file documenting ControlNet training loop, data format, JSON schema, configs, launch commands, and gotchas checklist.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Raw image + mask] --> B[Step 1: VAE encode]
    A --> C[Step 2: NV-Segment]
    C --> D[Step 3: Remap + combine]
    D --> E[Step 4: NN-resample to 4x latent]
    B --> F[JSON data list]
    E --> F
    F --> G[scripts.train_controlnet]
    G --> H[Checkpoints best + current]
    H --> I[infer_image-from-mask]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[Raw image + mask] --> B[Step 1: VAE encode]
    A --> C[Step 2: NV-Segment]
    C --> D[Step 3: Remap + combine]
    D --> E[Step 4: NN-resample to 4x latent]
    B --> F[JSON data list]
    E --> F
    F --> G[scripts.train_controlnet]
    G --> H[Checkpoints best + current]
    H --> I[infer_image-from-mask]

Loading

_{Reviews (4): Last reviewed commit: "data-prep: hand off to the train_control..." | Re-trigger Greptile}

[Can-Zhao]

@addsouza-nvidia May I ask if you can help review it? Thank you!

[addsouza-nvidia]

Addresses missing info about training data size by adding a new train_controlnet_image-from-mask skill, and extends the data-prep skill with Step 4 (resampling the combined label to the encoded-image grid).