FunPack Story Writer

FunPack Story Writer is the node that turns "a knight searching for a lost city" into five connected shot prompts, each describing one scene with characters, positioning, action, camera angle and motion. It's the plot department for the FunPack pipeline - the LLM-powered front end that feeds the shot lists which eventually become scenes in the Scene Chain Sampler or a stitch.

How it works

The flow has an extra layer most prompt nodes skip. In mode = Sequences from story (default), the model first writes a hidden story outline, then expands it into individual sequence prompts. In Sequences from user prompt, it skips the outline and generates shots directly from your request - faster, less structured, good for a single beat rather than a narrative. Either way, prompt_count (1–5) decides how many shots you get, and the outputs come out as prompt1..prompt5 (slots beyond prompt_count come back empty).

Then comes the part that makes it feel like a real assistant: sanity_check (default on) runs each generated sequence back through the same model against a sanity_check_system_prompt that rejects physically impossible actions and demands character/camera detail. It's a second LLM pass per shot, so it costs time, but it's what keeps "the knight flies" from sneaking into a prompt list the model can't honor.

Two controls shape continuity. disable_continuity (default off): when off, each later sequence is generated with the previous shot history fed back in, so the knight's appearance stays consistent across scenes - this is the closest thing this node has to a character-consistency mechanism. provide_current_id (default on) still tells the model which sequence number it's on when continuity is disabled. And vision_input is a text slot for "put the outputs of your VL model here" - paste a vision-language model's description of your starting image and the shots will stay aligned to it.

The inputs that matter

Same LLM plumbing as Prompt Enhancer: model_path_type (Local Safetensors / HuggingFace Pretrained), model_path (default mlabonne/NeuralLlama-3-8B-Instruct-abliterated), plus the sampler dials and an extra min_p (0.1). The per-shot prompt1..5 inputs let you pin instructions for specific slots while leaving others to the model.

And the same trap applies: this needs transformers >= 5.0, which the optional hpsv3 package breaks by pinning 4.45.2. Don't install both on one environment.

Install

It's part of ComfyUI-FunPack:

cd ComfyUI/custom_nodes
git clone https://github.com/olivv-cs/ComfyUI-FunPack
pip install -r requirements.txt

or via ComfyUI Manager (search "ComfyUI-FunPack"). First run downloads the 8B model.

Where it fits

Story Writer is the storyboard stage: its prompt outputs feed FunPack Prompt Combiner (to merge in a shared style base) or go straight into per-scene conditioning for the Scene Chain Sampler. The README's lorebook node pairs beautifully here too - inject worldbuilding first, then let Story Writer draft shots that respect it. It's opinionated and it makes assumptions, so read its output before committing a story to a 20-minute render.