🌸Flower Load Text From Folder

Here's the problem Flower Load Text From Folder was built to solve, and it's a real one: ComfyUI TTS pipelines like Qwen-TTS choke on ultra-long inputs - feed them a whole novel chapter as one text blob and you get an out-of-memory crash or a garbled read. The community fix people reach for first is "split your files into ≤1000-character chunks by hand," which is miserable. This node automates the whole thing: it reads .txt files from a folder, chops long text into segments at sensible boundaries (even chapter-aware), and hands you one segment per execution, driven by a seed - so the TTS reads piece by piece and never sees the whole book at once.

What it outputs

text is the payload. The other three outputs are the real convenience: full_path, file_name, and file_name_no_ext - strings describing which file (or virtual chapter segment) the current text came from. That's what lets you save each narrated segment under a matching name instead of guessing later.

The knobs that matter

• directory - where the .txt files live. filter_keyword / negativeKeyword prune the list, and the pattern language is worth learning: A&B|C means "(A AND B) OR C", and blank means no filtering.
• seed + continuous_processing - your position in the file list (or segment list). continuous_processing > 1 holds the same pick for that many consecutive images/runs.
• max_chars_per_segment - the core dial. -1 (default) reads whole files; set it to e.g. 500 and the node segments text and steps through segments seed-by-seed, cutting at punctuation (split_symbols) so it never chops mid-sentence.
• split_by_chapter - the dropdown with three modes. 無 (no chapter logic, just char-count segments), 根據章編號分章輸出(第一章) (detect 第X章 headings - Chinese or Arabic numerals, brackets tolerated - and keep each chapter's audio self-contained, default), and 根據字數自編章節輸出(4000字合成一章) (auto-chunk by chars_per_auto_chapter, cutting at line breaks, merging a short tail into the previous block). In chapter modes the reported file_name gets a label like 喜寶-V400-AI語音用-001第一章-0001-984, so saved audio files carry their chapter on the filename.

There's a Refresh button that runs the real backend logic without executing the graph - it updates a content preview and the virtual segment list, which is genuinely handy for checking "is my chapter detection working" before you burn GPU time.

Why it exists

The README is explicit: this node exists because Qwen-TTS crashed on long text, and manual splitting was the workaround. The intended pipeline is FlowerLoadTextFromFolder → FlowerSplitSentences → TTS, optionally with FlowerFileNameCombination naming the output audio. If you're narrating audiobooks or long-form voiceover in ComfyUI, this is the missing middle - it's the only node in the pack whose whole identity is "make long text survivable for TTS."

Install

cd ComfyUI/custom_nodes
git clone https://github.com/weichenglin0215/comfyui-flower-tools.git

Restart (or ComfyUI Manager → search ComfyUI Flower Tools). No models, no downloads.

Gotchas

• Segmentation only kicks in at max_chars_per_segment ≥ 256. Set it below that and the node falls back to whole-file reading. The exact threshold matters more than the README lets on.
• The text output changes per seed, but the graph only re-runs when the seed or settings change - that's normal ComfyUI caching, not a bug. Bump the seed to step through segments.
• full_path for a virtual chapter is virtual - it points at the source file with a synthetic chapter-suffixed name, not a real on-disk file. Fine for naming saves, but don't try to open that path directly.
• Encoding is assumed UTF-8; files in other encodings can read back as mojibake.