This node generates up to 5 sequential prompts for video shots based on a user request, optional per-shot instructions, and an LLM.
It is built around LLaMA-style chat models and is intended for story-driven multi-shot generation workflows.
user_prompt: Main instruction describing the overall story, scene, or video idea.
prompt1, prompt2, prompt3, prompt4, prompt5: Optional extra instructions for specific sequence slots. Leave them empty if you want the node to decide each shot from the main request alone.
story_system_prompt: System instruction used when the node first generates a hidden story outline.
sequence_system_prompt: System instruction used when generating the actual sequence prompts.
model_path_type: Choose between a local .safetensors LLM file and a HuggingFace pretrained model.
model_path: HuggingFace repository path used when model_path_type is set to pretrained.
llm_safetensors_file: Local .safetensors model file from your ComfyUI models/clip folder.
prompt_count: Number of sequence prompts to generate, from 1 to 5.
top_p, top_k, min_p, temperature, max_new_tokens, repetition_penalty: Text generation settings that affect variety, strictness, and output length.
mode: Sequences from story first creates a hidden story and then expands it into shots. Sequences from user prompt skips the hidden story step and generates shots directly from the user request.
vision_input: Optional text from a vision-language model describing the starting image so the generated shots can stay aligned with it.
sanity_check: If enabled, each generated sequence is reviewed by the same model and corrected if it breaks the requested rules.
sanity_check_system_prompt: System instruction used for the sanity-check pass.
disable_continuity: If enabled, later sequences are generated without feeding the previously generated shot history back into the model.
provide_current_id: When continuity is disabled, this controls whether the current sequence number is still provided to the model.
prompt1, prompt2, prompt3, prompt4, prompt5: Generated sequence prompts. Outputs beyond prompt_count are returned as empty strings.
Use this node when you want one LLM pass to turn a high-level idea into a short sequence of video prompts that still feel connected from shot to shot.