---

name: stylegan-human-manipulation description: "Guides StyleGAN-Human alignment, background preprocessing, PTI inversion, latent-direction editing, real-image workflows, InsetGAN, and asset preflight." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

StyleGAN-Human Manipulation

Use this route for human-image preparation and editing rather than basic generation: whitening backgrounds, aligning raw photos, inverting real images with PTI, applying upper_length/bottom_length edits, or composing face/body images with InsetGAN.

Asset-first workflow

Run the preflight helper before touching model code:

python sub-skills/stylegan-human-manipulation/scripts/check_stylegan_human_assets.py \
  --repo-root /path/to/DragGAN \
  --check-alignment --check-pti --check-editing --check-insetgan

Missing assets are expected in a fresh checkout. Do not replace a missing model, latent direction, or segmentation file with a random download.

Choose a route

• CPU-safe foreground/background processing: references/alignment-and-data-prep.md and the bundled bg_white.py.
• Raw human photo to aligned 512x1024 image: alignment prerequisites and the one-person constraint in the same reference.
• Real image to editable latent: PTI configuration and outputs in references/inversion-and-editing.md.
• Generated or PTI-edited human image: attribute editing with upper_length or bottom_length.
• Face/body fusion: InsetGAN, which needs separate face/body checkpoints and dlib models.

Background whitening

The bundled helper is self-contained and CPU-safe:

python sub-skills/stylegan-human-manipulation/scripts/bg_white.py \
  --raw-img-dir raw_images --raw-seg-dir masks --outdir white_background

It requires matching filenames and writes one output per valid raw/mask pair. Use a tiny fixture first; the helper returns non-zero when no pair can be processed.

Real-image route

For real-image editing, keep the order explicit: raw image -> alignment -> PTI/e4e inversion -> attribute edit. PTI path and global settings are file-based in the source workflow, so use the reference checklist and preflight helper before launching a long CUDA optimization.

Related routes

• Basic generation, interpolation, and style mixing: ../stylegan-generation/SKILL.md.
• Interactive DragGAN point editing: ../draggan-ui/SKILL.md.
• SHHQ/StyleGAN2/3 training: ../stylegan-training/SKILL.md.

Bundled files

• scripts/check_stylegan_human_assets.py checks expected assets without downloading.
• scripts/bg_white.py adapts the deterministic background-whitening utility.
• references/alignment-and-data-prep.md covers image/mask layouts and alignment prerequisites.
• references/inversion-and-editing.md covers PTI, edit, InsetGAN, and configuration.
• references/troubleshooting.md covers optional dependencies and failure recovery.