---

name: interactive-ui description: "Launch and operate the legacy PyQt4 iGAN interactive interface, drawing tools, candidate grid, slider, control panel, AverageExplorer, and ShadowDraw modes." disable-model-invocation: true metadata: disco-role: operating license: MIT

interactive-ui

Use this sub-skill when the task is to open, configure, explain, or troubleshoot iGAN_main.py, the PyQt4 interactive drawing interface for iGAN.

Read First

1. For launch recipes and preflight commands, read references/ui-workflows.md.
2. For what each visible widget, brush, slider, shortcut, and mode does, read references/controls-reference.md.
3. For concrete failure diagnosis, read references/troubleshooting.md.
4. For pretrained model files, sample generation smoke tests, and artifact URLs, route to ../model-inference/SKILL.md.
5. For headless color/mask/edge-map constraint generation, route to ../constraint-generation/SKILL.md.

Use When

• The user wants a dry command for launching the iGAN GUI.
• The task mentions PyQt4, drawing pad, candidate results, color/sketch/warp brushes, slider interpolation, Play/Fix/Restart/Save/Edits, AverageExplorer, or ShadowDraw.
• The user asks how to operate the interactive interface once it appears.
• The user asks why the UI fails to open, opens blank, cannot import PyQt4, cannot find a display, cannot load a model, or does not update candidates.
• The user needs an optional native verification candidate for an interactive launch but has not asked to run a GPU/UI session yet.

Do Not Use When

• The user only wants non-UI generation from color masks and edge maps; use ../constraint-generation/SKILL.md.
• The user asks to download DCGAN/AlexNet artifacts or generate sample grids; use ../model-inference/SKILL.md.
• The user asks to train DCGAN models, create HDF5 datasets, or pack training checkpoints; use ../training-data/SKILL.md.
• The user asks to project an image into latent space; use the projection sub-skill if present, not this UI router.
• The user asks to modernize or port the code to a new framework; this sub-skill can identify UI dependencies, but it does not own a rewrite plan.

Runtime Facts To Preserve

• iGAN_main.py is a legacy script that imports PyQt4, qdarkstyle, Theano model definitions, and a framework-specific constrained optimizer during startup.
• The documented runtime assumes Python2-era packages, Theano, OpenCV, CUDA, cuDNN, and a desktop display or VNC session.
• The default model command uses --model_name outdoor_64 and resolves the model file to ./models/<model_name>.<model_type> when --model_file is omitted.
• --model_type defaults to dcgan_theano; --framework defaults to theano and selects constrained_opt_theano through dynamic lookup.
• --top_k controls the maximum number of candidate thumbnails; --batch_size controls random initial latent samples used by optimization.
• --n_iters controls per-edit constrained optimization iterations; real-time behavior depends on a compatible GPU stack.
• --morph_steps controls slider/playback frames between the previous result and the current edited result.
• --average enables AverageExplorer display, where the main pad may show a weighted average of candidate results.
• --shadow enables ShadowDraw-style sketch assistance and disables color and warp buttons in favor of sketching.
• hed_shoes_64 is the documented model for ShadowDraw and should normally be combined with --shadow --average.

Build Launch Commands Safely

Use the bundled dry-run helper instead of importing or executing the GUI:

python sub-skills/interactive-ui/scripts/build_igan_command.py --help
python sub-skills/interactive-ui/scripts/build_igan_command.py \
  --model-name outdoor_64 --check-display

For ShadowDraw with AverageExplorer:

python sub-skills/interactive-ui/scripts/build_igan_command.py \
  --model-name hed_shoes_64 --shadow --average --check-display

The helper prints the THEANO_FLAGS environment and python iGAN_main.py ... command but never launches PyQt4, never loads a model, never downloads files, and never touches CUDA.

Launch Readiness Checklist

Before telling the user to run the command, confirm these facts from references/ui-workflows.md:

• A compatible Python/PyQt4/OpenCV/qdarkstyle stack is active.
• The requested pretrained model file exists under the expected models/ path or the user supplied --model_file.
• The display is available through local X11, VNC, Xpra, or another remote desktop path; a bare SSH shell without DISPLAY is not enough.
• THEANO_FLAGS selects the intended device, usually device=gpu0,floatX=float32 for the legacy real-time path.
• The user understands that modern Python and modern CUDA may need pinning, compatibility shims, or a container built for the old Theano stack.

Operate The UI

• Explain the window as a drawing pad on the left and a candidate grid on the right; details live in references/controls-reference.md.
• For color edits, choose Coloring, right-click the color chip or pad to select a color, drag with left mouse, and use the wheel to change brush width.
• For shape edits, choose Sketching and drag with left mouse; in ShadowDraw mode this is the primary enabled tool and color toggles between black and white.
• For warping, first create color/sketch constraints, right-click a square source patch, then left-drag the patch target; use the wheel for patch size.
• Candidate thumbnails update during constrained optimization; click a thumbnail to select it for the main drawing pad.
• The slider and Play button inspect morph frames after an edit; Fix promotes the current image/frame as the next latent starting point.
• Restart clears the UI state; Save writes an HTML/image result in the user-chosen output folder after the first save dialog.
• Keyboard shortcuts include P, F, R, S, E, A, and Q.

Troubleshooting Routing

• If import fails before a window appears, start with dependency and Python version rows in references/troubleshooting.md.
• If a Qt error says no platform/display is available, diagnose display/VNC setup before touching Theano or model files.
• If the window opens but candidates never update, inspect CUDA/Theano/model loading, the constrained_opt_theano dynamic import, and optimizer logs.
• If ShadowDraw is requested with a non-sketch model, recommend hed_shoes_64 or explain that other models may not provide useful sketch guidance.
• If the user wants to bypass the UI after display problems, route to ../constraint-generation/SKILL.md.

Native Verification Candidate

Preserve the optional native case as a blocked/unverified runtime launch unless all prerequisites are available:

THEANO_FLAGS='device=gpu0,floatX=float32,nvcc.fastmath=True' \
python iGAN_main.py --model_name outdoor_64

Expected signal: a PyQt4 window titled Interactive GAN opens, model arguments print to stdout, the candidate grid updates after a stroke, and quitting with Q exits the application. Do not run this case in an automated verification session without explicit UI, model-file, and GPU approval.