cohub-generate

# Cohub Multimodal Generation Use `cohub generate` to create or transform images, video, speech, and music. Prefer simple, explicit commands. Use `--json` when reading output for decisions, extracting URLs, or chaining commands. If a target Space is required, add `-s "$COHUB_SPACE_ID"` or an explicit Space ID. ## Installation If `cohub` is unavailable, install it and check availability: ```bash npm install -g @neta-art/cohub-cli cohub --help ``` ## Models Always resolve models from the live list; do not hardcode a catalog. ```bash cohub models ls --model-type multimodal --json ``` Typical categories: text-to-image / image editing, text-to-video / image-to-video, text-to-speech (TTS), background removal, music. Before using non-default parameters or reference media, inspect the model schema for supported inputs, roles, parameters, defaults, and examples: ```bash cohub models show <model> --json ``` ## Generate Default to returning generated result links directly, with Markdown preview when possible (`` for images, direct links for video and audio). Save locally only when the user asks for a file download, local editing, or post-processing. Generate from text: ```bash cohub generate "a calm lake at sunrise" \ --model <model> ``` Generate with reference media: ```bash cohub generate "restyle this image" \ --model <model> \ --image ./input.png \ --param size=1024x1024 ``` Supported inputs: `--image`, `--video`, and `--audio`, each repeatable. Pass a URL or a local path; local files upload to an unlisted public URL first, so tasks store a reference instead of inline data. For files that must stay private, add `--inline` to keep them inside the task. When a model requires input roles, prefix the path or URL: ```bash cohub generate "smooth transition between two shots" \ --model <model> \ --image first_frame=https://example.com/first.png \ --image last_frame=https://example.com/last.png cohub generate "keep these characters consistent" \ --model <model> \ --image reference_image=https://example.com/a.png \ --image reference_image=https://example.com/b.png cohub generate "lip-sync to this spoken take" \ --model <model> \ --image reference_image=https://example.com/portrait.png \ --audio reference_audio=https://example.com/speech.mp3 ``` Roles include `first_frame`, `last_frame`, `reference_image`, `reference_video`, and `reference_audio`. Check `models show` for what a model accepts. Do not mix first/last frame roles with reference roles. Seedance 2 reference audio needs an image or video in the same request. Pass generation parameters with `--param key=value` (repeatable; JSON, number, or boolean values) or `--parameters '<json>'`: ```bash cohub generate "cinematic drone shot over misty mountains" \ --model <model> \ --param duration=5 \ --param resolution=720p \ --param ratio=16:9 ``` Results print their media facts when available, and videos their last frame: ```text video 720×1280 · 11.0s: https://…/clip.mp4 last frame: https://…/last.webp ``` With `--json`, the same facts are in `outputMedia` (`index` matches `output`). Other useful flags: ```bash cohub generate "..." --model <model> --output ./out.png # save locally cohub generate "..." --model <model> --async # queue and return cohub generate "..." --model <model> --timeout-ms 120000 # sync wait limit cohub generate "..." --model <model> --meta '<json>' # pass model metadata ``` ## Speech (TTS) Design a voice from text: ```bash cohub generate "Welcome to Cohub. This voice was created from a description." \ --model qwen-audio-3.0-tts-plus \ --meta '{"voice_prompt":"A calm, clear male narrator with a warm tone"}' ``` Clone one voice from a public reference URL: ```bash cohub generate "Welcome to Cohub. This voice follows the reference recording." \ --model higgs-tts \ --audio "$REFERENCE_AUDIO_URL" ``` ## Workflows Animate a still image: ```bash cohub generate "<motion prompt>" \ --model <model> \ --image first_frame=./still.png ``` Continue a clip: pass its last frame as the next clip's first frame: ```bash cohub generate "<next shot>" \ --model <model> \ --image first_frame=<last frame URL> ``` Edit or restyle with a reference image: ```bash cohub generate "<edit instruction>" \ --model <model> \ --image ./input.png ``` For less common options, use `cohub generate -h`.