VRGDG_LLM_OutputSaver

Long music-video runs have a problem: the LLM generates prompts for dozens of scenes, the pipeline uses them, and then the run is over and all that creative work is gone. VRGDG_LLM_OutputSaver is the pack's answer - it appends each batch of LLM output to a text file on disk so you've got a permanent record (and a resumable one). If a good run happens, you can re-read the prompts and understand exactly what produced it.

This matters more than it sounds. Community music-video runs with this workflow are genuinely long - one user reported ~17 hours of generation on a 12GB 4070 Ti, and people routinely iterate on the same track across days. When a run breaks at hour nine, or you want to re-render one scene with a tweaked setting, the saved prompts are your diff tool.

What it does

• text (STRING, multiline) - the LLM output to save, wired from something like VRGDG_LLM_Multi.
• batch_index (INT) - which batch this chunk is. The node uses it to append in order, not overwrite.
• is_final_batch (BOOLEAN) - flip this on the last batch to signal the write is complete.
• output_folder (STRING) - where files land.
• base_filename (STRING, default LLM_Output) - the filename prefix, so consecutive runs don't clobber each other.
• combined_text (STRING, output) - the accumulated text across all batches, handy to forward downstream.

It's marked as an output node, meaning it's designed to sit at the end of a branch and just do its job. The mechanism is straightforward: each batch appends its text to the current file, batch_index keeps the order sane, and combined_text gives you the whole thing back in one string if you want to feed it onward.

Where it fits

In the chunked pipeline, the LLM doesn't write one giant prompt - it writes per-scene or per-set prompts, batch by batch, as the queue advances. Saver sits at the end of that branch. When the run finishes you have, say, LLM_Output_001.txt full of every scene instruction, in order. VRGDG_LoadTextAdvanced in this pack can then read those files back on a later run - that's the loop that makes iteration possible instead of starting from scratch.

Installing it

Part of the comfyui-vrgamedevgirl pack. ComfyUI Manager: search "vrgamedev", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

then install the README's requirements (kornia, librosa, imageio):

pip install -r custom_nodes/comfyui-vrgamedevgirl/requirements.txt

If it's not working

The classic mistake is leaving is_final_batch false forever, then wondering why the file looks incomplete - the node is waiting for a final signal before it considers the batch closed. Also: the file is only as good as the folder you give it. Point output_folder somewhere you'll actually find again, and keep base_filename unique per run or you'll silently overwrite the previous run's record. If you're rerunning and the earlier text vanished, that's the usual culprit.