---

name: cross-modal-generation description: "Guides InternGPT ImageBind/StableUnCLIP audio, thermal, image, text generation and StyleGAN/DragGAN point editing workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Cross-Modal Generation

Use this sub-skill when a task asks for InternGPT generation or editing through ImageBind, StableUnCLIP, StyleGAN, or DragGAN: audio-to-image, thermal-to-image, audio+image-to-image, audio+text-to-image, "ImageBind", "StableUnCLIP", "DragGAN", "StyleGAN", "New Image", "click start/end points", or "Drag It".

Route first

• For ImageBind/StableUnCLIP workflows, read references/imagebind-workflows.md when you need tool names, input grammar, modality preprocessing, output naming, or sample-asset expectations.
• For StyleGAN and DragGAN point editing, read references/draggan-workflows.md when you need the app button sequence, state fields, point pairing rules, iteration behavior, or output artifacts.
• For common failures, read references/troubleshooting.md when an ImageBind wrapper is missing, a comma input is malformed, assets are absent, CUDA/model downloads fail, VRAM is insufficient, or DragGAN points/checkpoints fail.
• Run scripts/validate_multimodal_assets.py before invoking a multimodal tool when you need a safe local check for audio/image/thermal/video paths, extensions, and ImageBind comma-separated tool input grammar; it performs no model imports or downloads.

Boundaries

• This sub-skill owns Anything2Image, Audio2Image, Thermal2Image, AudioImage2Image, AudioText2Image, StyleGAN, and DragGAN app-point workflows.
• Route general launch flags, model-zoo placement, Docker, HTTPS, OpenAI key, and broad --load planning to ../app-deployment/SKILL.md.
• Route SAM, OCR, Husky, mask/inpainting, ControlNet, and general visual dialogue workflows to ../visual-dialogue-tools/SKILL.md.
• Route video upload, video caption/action/dense caption, TikTok clip creation, ffmpeg, and Bark workflows to ../video-understanding/SKILL.md.

Operating cautions

Do not treat static validation as proof that full generation will run: actual ImageBind/StableUnCLIP and StyleGAN/DragGAN inference requires the runtime package stack, CUDA-capable PyTorch, large model checkpoints or model downloads, enough VRAM, and a configured app/service session. Prefer safe validators and references until the user explicitly asks to run the heavy runtime and confirms the environment is ready.