Add train_controlnet_image-from-mask skill - #40

Merged

Can-Zhao merged 15 commits into

NVIDIA-Medtech:mainfrom

Can-Zhao:skill/train-controlnet-image-from-mask

Jul 16, 2026

Commits on Jul 15, 2026

Add train_controlnet_image-from-mask skill

Can-Zhao
and
claude
committed
01320b6
Address review: valid JSON defaults, clearer label-grid wording

Can-Zhao
and
claude
committed
3777a4c
Apply review: correct AE usage, modality mapping, orientation caveat

Can-Zhao
and
claude
committed
b9d7240
data-prep: require the combined label on the encoded-image grid

Can-Zhao
and
claude
committed
3c311df
data-prep: promote label-grid resampling to a concise Step 4

Can-Zhao
and
claude
committed
bdf333a
data-prep: spell out the label resample target in the Step 4 example

Can-Zhao
and
claude
committed
91d1efb
data-prep: boil Step 4 down to the core rule (label = 4x latent per axis)

Can-Zhao
and
claude
committed
3cc5104
data-prep: drop "In short" lead-in from Step 4

Can-Zhao
and
claude
committed
995550d
train skill: shorten Spatial relationship, defer resample how-to to data-prep Step 4

Can-Zhao
and
claude
committed
479cae0
train skill: drop resample detail from Spatial relationship, link only

Can-Zhao
and
claude
committed
3d7f957
train skill: promote label/embedding orientation caveat to its own subsection

Can-Zhao
and
claude
committed
d121055
train skill: explain why Orientationd is label-only (embeddings already RAS)

Can-Zhao
and
claude
committed
d0c9eb4
train skill: note adding "image" to Orientationd is a harmless hardening

Can-Zhao
and
claude
committed
32016ae
Orient image embedding to RAS in the controlnet dataloader

Can-Zhao
and
claude
committed
db3840d
data-prep: hand off to the train_controlnet_image-from-mask skill

Can-Zhao
and
claude
committed
64909d0

Navigation Menu

Uh oh!

Add train_controlnet_image-from-mask skill - #40

Add train_controlnet_image-from-mask skill#40
Can-Zhao merged 15 commits into
NVIDIA-Medtech:mainfrom
Can-Zhao:skill/train-controlnet-image-from-mask

Commits on Jul 15, 2026

Add train_controlnet_image-from-mask skill

Address review: valid JSON defaults, clearer label-grid wording

Apply review: correct AE usage, modality mapping, orientation caveat

data-prep: require the combined label on the encoded-image grid

data-prep: promote label-grid resampling to a concise Step 4

data-prep: spell out the label resample target in the Step 4 example

data-prep: boil Step 4 down to the core rule (label = 4x latent per axis)

data-prep: drop "In short" lead-in from Step 4

train skill: shorten Spatial relationship, defer resample how-to to data-prep Step 4

train skill: drop resample detail from Spatial relationship, link only

train skill: promote label/embedding orientation caveat to its own subsection

train skill: explain why Orientationd is label-only (embeddings already RAS)

train skill: note adding "image" to Orientationd is a harmless hardening

Orient image embedding to RAS in the controlnet dataloader

data-prep: hand off to the train_controlnet_image-from-mask skill