LTX-GapFill Inspector

The GapFillInspector (shown as LTX-GapFill Inspector) is the boring node in this pack, and that's exactly why you want it. Its sibling GeminiFillPrompt sends your frames and prompts to Google's Gemini API - which costs you quota and a couple of seconds per call. This node shows you the entire payload that would be sent, before a single byte goes over the wire. It's a passthrough debug node: it sits in-line ahead of the main node, builds a full report of the model, system prompt, frame sizes, and estimated payload - then passes every input through unchanged.

Think of it as the "dry run" button for prompt engineering. LTX already rewards long, careful prompts (the model is famously sensitive to prompt wording), and this pack's whole job is generating those for bridge shots. The Inspector is where you catch a mistake before it costs you a wasted generation and a chunk of your daily quota.

How it works

There's no API key input and no network call - it's pure local math that mirrors the main node's logic. It assembles the exact system prompt Gemini would receive (based on your prompt_style, with the per-style token budget), resizes your frames the same way the real node would, and computes what gets sent: original resolution, resized resolution, JPEG size, and an estimated total payload in KB. The resulting report is a neatly boxed ASCII table covering model, endpoint, style, gap duration, resize settings, and the full system prompt text - plus per-frame info and prompt previews for each side of the gap.

One detail that justifies the node on its own: when you feed it a full video batch, the report notes batch=N, using last/first, so you can verify the right frame is being auto-selected before you commit to a generation.

Inputs that matter

• prompt_style - pick the style whose system prompt you want to preview. This is the whole point: you see the actual text Gemini will get.
• gap_duration, resize_before_send, max_size_px - mirror the main node and affect the size estimate, which matters because Gemini's inline payload cap is ~20MB.
• custom_system_prompt - drop a custom system prompt here and the report shows it, flagged with a warning that it overrides prompt_style entirely. Great for testing your own prompt before spending tokens on it.
• print_to_console - defaults to off; wire the report output into a Show Text node instead.

It has seven outputs, and only two are interesting: report (the payload preview, for Show Text) and active_model (the resolved model name - wire it into the main node's custom_model to keep both in sync). The other five - prompt_style, prompt_before, prompt_after, clip_before_last_frame, clip_after_first_frame - are straight passthroughs so the node can sit in-line without breaking your connections.

Setting it up

It ships in the same pack as GeminiFillPrompt, so one install gets you both. Install via ComfyUI Manager (search "ComfyUI-LTX-GapFill") or:

cd ComfyUI/custom_nodes
git clone https://github.com/PixWizardry/ComfyUI-LTX-GapFill

Restart ComfyUI and both nodes appear under LTX → Gap Fill. There are no model downloads and no extra dependencies - requests, Pillow, and numpy, which your ComfyUI already has. The Inspector doesn't even need the Gemini API key, since it never calls the API.

Gotchas

Because it's marked as an output node, it always executes even if its outputs aren't consumed - which is what you want from a debug node, but it means it fires on every run whether you're looking or not. Two things to remember: wire report into Show Text (the console default is off), and don't expect it to validate your API key or catch quota problems - it can't, since it never makes the request. It's a seeing-eye dog, not a security guard. If the payload looks right, the main node is the one that'll tell you whether Google agreed.

It's a small, single-purpose tool from a small author, and it earns its place. Most packs make you learn what you're sending the hard way; this one shows you first.