MultiTrack Prompt Enhance To Project

A multi-shot MiniMax H3 project in the Easy Media MultiTrack Editor is a list of task segments, each with its own prompt. The fiddly part of long-form work is keeping every segment's prompt rich and consistent - which is exactly the job an LLM prompt enhancer is good at, and the job you'd normally do segment by segment, copying text between a half-dozen nodes. MultiTrack Prompt Enhance To Project is the batch version: it walks every task in your TRACKS_INFO, runs each prompt through an LLM, writes the enhanced result back into its own segment, and hands you project data that's ready to feed straight into MultiTrack Project.

It sits between the editor and the generator:

MultiTrack Editor --TRACKS_INFO--> multiTrackPromptEnhanceToProject --TRACKS_INFO--> MultiTrack Project

You'd reach for this over the per-segment easy multiTrackPromptEnhancer when you want the whole timeline treated as one pass - consistent style, one seed, one setting - instead of babysitting each shot.

How it works

Under the hood it's text-only inference, one call per task segment, run synchronously. For each segment it reads the prompt (honoring your A/B variant choice), the task type, and the length, then builds a request the provider understands - and note it deliberately does not upload each segment's reference images or video. The enhanced text is then written back into that task segment's user-prompt field via the same code path the Apply node uses, and both the updated TRACKS_INFO and the plain PROMPTS list come out the other end.

Which model does the enhancing? That's the model dropdown, and it's the real story here - the pack supports a roster rather than one vendor:

• MiniMax h3-context-ir (the default), MiniMax's official H3 context API
• OpenAI-compatible third-party multimodal LLMs - doubao-seed and GLM via Volcano Engine, Zhipu, or RunningHub
• local llama.cpp, if you'd rather not pay and don't mind slower per-segment calls

API keys come from config.yaml in the pack folder or environment variables (MINIMAX_API_KEY, VOLCENGINE_API_KEY, BIGMODEL_API_KEY, RUNNINGHUB_API_KEY). Selecting a provider from the combo reveals its sub-options - the API-key box, token caps, ratio for MiniMax. That api_account widget at the bottom is not an input you fill in: it's a balance and key-status readout for the providers, and per the tooltip it has no effect on the enhancement itself.

Inputs and outputs

The three you actually touch: tracks_info (in), model (which provider), and enabled. The seed widget sets one generation seed reused for every segment, and llama_model only matters when you've picked the local llama.cpp option - it needs the llama_cpp_instruct_adv node from ComfyUI-llama-cpp_vlm. Both outputs - TRACKS_INFO (prompts now embedded in the segments) and PROMPTS (the ordered list) - plug into MultiTrack Project or downstream nodes.

Installing, costs, and the honest caveats

Install the pack through ComfyUI Manager ("ComfyUI-Easy-Media") or git clone https://github.com/yolain/ComfyUI-Easy-Media into custom_nodes, add FFmpeg if you haven't, restart. Then the first real step is putting at least one key in config.yaml, because with enabled on and no key, every segment call just fails.

Two things worth knowing before you wire it in. First, this is a metered node: every enabled segment is a billed API call, and a 30-shot project is 30 calls. Local llama.cpp avoids the meter but trades it for VRAM time. Second, your prompts (and only your prompts - it's text-only) leave the machine for whichever provider you chose. Easy Media is a legit pack from yolain, the same author behind the widely-installed ComfyUI-Easy-Use, so this isn't a warning to avoid it - it's context for deciding whether the local model is worth the setup. Text-only also means the enhancer can't see your reference images: great for cost and speed, just don't expect it to describe the look of a shot it can't see.