Rewrite each node to faithfully mirror the official comfy_api_nodes/
contract (apis/*.py + Comfy-Org/embedded-docs). The ComfyOrg proxy
forwards the body verbatim, so the official request/response models
ARE the real vendor API.
- Video nodes now return VIDEO (InputImpl.VideoFromFile), not file paths
- Audio nodes now return AUDIO ({waveform, sample_rate}), not file paths
- Request/response bodies match official pydantic models field-for-field
- Model ids, options, defaults, ranges match official + embedded docs
- Every node exposes api_key + base_url (overridable for relays/proxies)
- utils.py: add ComfyUI-native helpers — video_from_bytes,
download_video_output, audio_from_bytes, download_audio_output,
poll_until, validate_prompt, tensor_to_data_uri
- Rewrite nodes_minimax.py (submit -> poll -> retrieve -> download VIDEO)
- nodes_extra.py: base_url on all three, poll_until, validate_prompt
- Clean __init__.py (drop duplicate grok import/keys); 37 nodes total
- README: output-type alignment, node list, relay examples
Co-Authored-By: Claude <noreply@anthropic.com>
Lightweight custom nodes that call AI APIs directly with your own keys,
bypassing the ComfyOrg proxy. Every node has a configurable base_url
for third-party relay/proxy support.
Covers: OpenAI, Claude, Gemini, Flux, Stability AI, Grok, Ideogram,
Recraft, Reve, Bria, Kling, MiniMax, Seedance 2.0, Runway, Luma,
Veo, Wan, PixVerse, Vidu, LTXV, ElevenLabs, OpenRouter, ByteDance,
WaveSpeed, Magnific, and a generic video gen node.
