Improper usage of PNG tEXt chunks in `parameters`

[enslo]

What happened?

Non-Latin-1 characters in the "parameters" metadata are being stored directly within a tEXt chunk. This violates the PNG specification (ISO-8859-1) and results in character corruption (mojibake) when the file is read by standard-compliant parsers.

Expected

When the parameters chunk contains non-Latin-1 characters:

• The parameters should be stored in an iTXt chunk.
• OR
• Any non-Latin-1 characters should be Unicode-escaped if they are kept within a tEXt chunk.

Bug

• Non-Latin-1 characters are stored directly as raw bytes in the tEXt chunk without escaping.
  • This violates the PNG specification, which requires tEXt chunk content to be ISO-8859-1 (Latin-1).
  • As a result, character corruption (mojibake) occurs in standard-compliant parsers (e.g., テスト becomes ���).

Steps to reproduce

1. Generate an image using a prompt that contains non-Latin-1 characters (e.g., テスト).
2. Inspect the parameters chunk of the generated PNG file.

Sidenote

• ComfyUI addresses this by using Unicode escaping while maintaining the use of tEXt chunks.
• Most other tools (especially A1111 derivatives, for compatibility with the parameters chunk content) switch from tEXt to iTXt chunks when the content includes non-Latin-1 characters.

I noticed this while developing my @enslo/sd-metadata library. Technically, this might be a minor point, but I wanted to let you know for better specification compliance.

Version

2.16.0-dev.1+597a23f

What Operating System are you using?

Windows