Simple_Readable_Metadata_Save_Prompt (NEGATIVE)-SG

The negative prompt is the half of the recipe people actually forget to save. In ComfyUI, if your negative CLIP Text Encode node's text box is fed by a wire instead of typed-in text, the embedded workflow records the link, not the words - so anyone reading the metadata later sees (empty) for the negative prompt. SaveNegativePromptSG is the mirror of SavePositivePromptSG: it guarantees your real negative text lands in the saved image's metadata.

It's a passthrough with a side effect. You wire your final negative text into its text input, and it flows straight out the connect_to_CLIP_text_encode output into your negative CLIP Text Encode node - your generation is untouched. While it runs, the node quietly rewrites its own entry in the workflow's prompt dict with the literal string and stuffs the text into extra_pnginfo under a NegativePrompt_<node-id> key. The SaveImage node embeds that, so the PNG carries your actual negative prompt, readable by the pack's metadata nodes.

The input and output

Two fields, that's the whole node:

• text (STRING) - "FINAL TEXT to be connected to CLIP Text Encode Negative node's box," per the author's tooltip. This is where your negative-prompt construction chain ends.
• connect_to_CLIP_text_encode (STRING) - wire this into your negative CLIP Text Encode node.

No settings, no toggles, nothing to tune. If you use the positive version on the other box, the two sit as a matched pair around your sampler.

Installing

Same pack, same one-time setup:

cd ComfyUI/custom_nodes
git clone https://github.com/ShammiG/ComfyUI-Simple_Readable_Metadata-SG.git
cd ComfyUI-Simple_Readable_Metadata-SG
pip install -r requirements.txt

Or Manager → Custom Nodes Manager → search Simple Readable Metadata-SG → install → restart. No model downloads.

Where it helps and where it can't

This node is for future generations. It injects the text at save time, so it does nothing for images you already have on disk - for those, use the pack's Text Viewer to dump all embedded workflow text and hunt down the negative by hand. And like everything that relies on embedded metadata, it's useless if a host re-encodes the image and strips the chunks. One more honest note from the README: when you later copy a prompt back out of a saved image, verify the text - the node guarantees storage, not that you'll recognize what comes back.

On ComfyUI 0.3.76+, remember the pack-wide catch: keep Node 2.0 switched off or these nodes' display widgets won't behave. The saving logic itself is plain Python and survives fine; it's the frontend rendering that breaks.

If you're wondering why the prompt was ever empty in the first place, read the positive version's page - it's the same quirk, and understanding it makes both nodes make sense.