ComfyUI Node
H3 Full-Reference Video Prompt Creator
Multimodal H3 full-reference prompt creator with IMAGE, VIDEO and AUDIO inputs, Qwen3-VL reasoning, faster-whisper transcription and librosa/FFmpeg audio analysis.
H3 Full-Reference Video Prompt Creator
- reference_video
- reference_audio
- reference_image_1
- reference_image_2
- reference_image_3
- reference_image_4
- reference_image_5
- reference_image_6
- h3_prompt
- reference_analysis
- audio_analysis
- generation_notes
◄idea►
◄target_duration6.0►
◄reference_intentAuto►
◄reference_notes►
◄audio_transcript_or_notes►
◄providerOllama (Local)►
◄api_key►
◄api_model►
◄ollama_urlhttp://127.0.0.1:11434►
◄ollama_modelqwen3-vl:8b►
◄temperature0.25►
◄keep_alive10m►
◄request_timeout600►
◄max_output_tokens4096►
◄num_ctx8192►
◄enable_reference_analysistrue►
◄whisper_modelsmall►
◄whisper_deviceauto►
◄enable_audio_transcriptiontrue►
CategoryH3 / Prompt Creator
Inputs (27)
| Name | Type | Default | Description |
|---|---|---|---|
| idea | STRING | Simple target idea. The node uses it together with any connected reference images, video and audio to write the complete six-section H3 full-reference prompt. | |
| target_duration | FLOAT | 6.00.1–120 | Target duration used to shape the generated timeline. If reference video duration is available, Ollama also receives it as evidence. |
| reference_videoopt | VIDEO | Optional full reference video. Sampled frames are sent to Ollama; the source is represented as <Video 1> in the H3 prompt. | |
| reference_audioopt | AUDIO | Optional reference audio. Local faster-whisper transcribes speech/lyrics and librosa extracts audio/music features; exact user notes can override ambiguity. | |
| reference_image_1opt | IMAGE | Optional picture reference 1. Automatically represented as <Picture 1> unless Ollama determines another H3 reference role. | |
| reference_image_2opt | IMAGE | Optional picture reference 2. | |
| reference_image_3opt | IMAGE | Optional picture reference 3. | |
| reference_image_4opt | IMAGE | Optional picture reference 4. | |
| reference_image_5opt | IMAGE | Optional picture reference 5. | |
| reference_image_6opt | IMAGE | Optional picture reference 6. | |
| reference_intentopt | COMBO | Auto | Optional hint. Auto is recommended: the node infers the task type from the connected references and notes. |
| reference_notesopt | STRING | Optional clarification. Leave blank when the references and simple idea are self-explanatory. | |
| audio_transcript_or_notesopt | STRING | Ollama itself is not assumed to hear arbitrary AUDIO tensors. Put exact lyrics/dialogue here when they must be preserved verbatim. | |
| provideropt | COMBO | Ollama (Local) | Ollama (Local) runs on your machine and needs no key. OpenAI, Anthropic, OpenRouter and Google Gemini are hosted APIs and need api_key + api_model. Built-in deterministic skips the model entirely. |
| api_keyopt | STRING | Key for the hosted providers. Leave blank to read OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY or GEMINI_API_KEY from the environment — safer, because a key typed here is saved into the workflow JSON and travels with it if you share the workflow. | |
| api_modelopt | STRING | Model for the hosted providers; ignored by Ollama. Blank uses the default: OpenAI gpt-4o, Anthropic claude-opus-5, OpenRouter anthropic/claude-sonnet-5, Gemini gemini-2.0-flash. Must be a vision model when you connect reference images. | |
| ollama_urlopt | STRING | http://127.0.0.1:11434 | Ollama server base URL. |
| ollama_modelopt | COMBO | qwen3-vl:8b | Vision-capable local model, listed from your running Ollama server. Use the Refresh Ollama Models button after pulling a new model. |
| temperatureopt | FLOAT | 0.250–1.2 | Lower values improve H3 schema adherence. |
| keep_aliveopt | COMBO | 10m | Keep the local Ollama model loaded between requests. |
| request_timeoutopt | INT | 60030–3600 | Maximum Ollama request time in seconds. |
| max_output_tokensopt | INT | 4096256–12000 | Maximum generated tokens. H3 full-reference descriptions benefit from 4096+. |
| num_ctxopt | COMBO | 8192 | Ollama context window. CRITICAL for speed: left unset, Ollama sizes the context from the model maximum (262144 for Qwen3-VL), which needs ~25 GB of KV cache and pushes most layers onto the CPU. 8192 keeps an 8B model fully on a 12 GB GPU. Raise it only when you connect many reference images. |
| enable_reference_analysisopt | BOOLEAN | true | Run the separate reference-dossier pass before writing the prompt. Turn off to halve generation time when references are simple. |
| whisper_modelopt | COMBO | small | Local faster-whisper model used for dialogue/lyrics transcription. Small is the recommended quality/speed balance. |
| whisper_deviceopt | COMBO | auto | Device for faster-whisper. |
| enable_audio_transcriptionopt | BOOLEAN | true | Run faster-whisper when audio is connected (or embedded in the reference video). |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| h3_prompt | STRING | — |
| reference_analysis | STRING | — |
| audio_analysis | STRING | — |
| generation_notes | STRING | — |