text_filter

Video prompting has a rhythm problem: you want "a girl walking under an umbrella" for frames 1–30, "rain starting, wind" for 31–60, and a completely different scene after that. The clean way to express that is a single text block with per-segment markers, and a node that picks the right segment for the current frame. text_filter is that node. It's the pack's per-frame prompt scheduler: give it one multiline text, a filter rule, and the current frame number, and it returns the segment that should be active at that moment - plus the full list of matches so you can see what it's choosing between.

The author's framing: "text filter preset" - and the presets are exactly the marker styles you've probably already been typing. @text@, "text", 'text', (text), {text}, text @, @text, or a custom rule. It wraps the regex so you don't have to.

What it actually does

Normalizes your text, builds a regex from the filter rule, extracts all matches, and then:

• match_all = off (default) → returns the first match... unless you use the trick below.
• match_all = on → joins all matches with newlines into one output.
• reverse_filter = on → inverts the filter, returning everything outside the markers.

The frame-scheduling magic is in the current_frame input: when match_all is off, the node cycles through the matches with frame_index % len(match_list). So a single text block like [sunny] cloudy stormy rain with rule [text] gives you a different segment per frame automatically. That's your per-frame prompt sequence with one node.

The inputs that matter

• text_input - required, multiline. Your marked-up prompt block.
• filter_rule - the marker style: None (no filter), @text@, @text, text @, "text", 'text', {text}, (text), or custom.
• custom_rule - when custom, you define the brackets: [text] extracts inside, [text extracts after, text] extracts before.
• current_frame - the frame counter from your video pipeline. This is what turns it into a scheduler.
• match_all - return every match joined, or single-select.
• reverse_filter - return the non-matching text instead.

The outputs that matter

调度文本 (scheduled text) - the string for the current frame, ready to feed a CLIP encoder or text_converter. Matched_list - the full list of extracted segments, so you can debug what the node is cycling through.

Wiring it in

For video: feed it the frame number (most video workflows expose one), take 调度文本 → text_converter → sum_TextEncode → stack → sampler. For static prompts it works the same but the frame stays 0.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git
cd ComfyUI-Apt_Preset
pip install -r requirements.txt   # Windows: double-click install.bat

Or ComfyUI Manager → ComfyUI-Apt_Preset.

Common issues

The failure modes are all about marker discipline. If your matches come back empty, the marker characters aren't in the text - note the rule names are literal templates ("text" means double-quote-delimited, not the word "text"). And empty or fully-filtered input returns the cheerful ❌ No valid content after filtering string, which will happily get encoded as a prompt if you don't watch for it - if your video suddenly goes wonky mid-run, check whether a segment got filtered to nothing. One more: current_frame with match_all off wraps around (modulo), so frame 999 on a 3-item list silently repeats item one - fine for loops, surprising for anything else.