DIGIT SRT From Video

If you've ever transcribed a video, you know the usual path is Whisper or a web tool, then a lot of manual SRT cleanup. DIGIT SRT From Video skips most of that: point it at a video file, and it extracts the audio, sends it to Gemini for transcription with timestamps, runs a proper subtitle post-processing pipeline, and hands you broadcast-ready SRT (or VTT, ASS, TXT - or burned straight into the video). It's from the same DIGIT pack as SRT Maker, but where Maker builds captions from a script, this one works from the audio. Two different jobs, complementary.

The appeal is the post-processing. Raw LLM transcription is messy - repeated hallucinated lines, 200-character entries, timestamps that land mid-frame. This node fixes those before you ever see the file: hallucination removal, line-length enforcement, frame padding, and snapping timestamps to frame boundaries. That last one matters more than it sounds on frame-accurate systems like Flame, where a subtitle that flickers between frames is a delivery failure.

The inputs that matter

• video_path - the file. Browse button in the UI, or paste a path.
• model - gemini-2.5-flash by default; you can step up to Pro or the Gemini 3 previews if accuracy on hard audio matters more than speed.
• subtitle_output - srt_only (sidecar files), burn_in_only (hardcode subs into the video), or both.
• extra_instructions - free text like "ignore background music" or "this is a commercial". Genuinely worth filling in; it's cheap accuracy.
• language - auto works, but specifying the language improves timestamps. translate_to adds a translation pass that preserves timing.
• The post-processing knobs: max_chars_per_line (42 = Netflix/broadcast standard), max_lines (2), pad_frames (breathing room), snap_to_frames, remove_hallucinations (on by default).
• Burn-in styling, only when you're hardcoding: font_name, font_size, font_color, outline_width, shadow_depth, position, margin_v. If you're just saving sidecars, ignore all of them.

Outputs: srt_filepath (where it saved) and srt_text (the raw content, handy for piping into SRT Preview or SRT Tools for more QA).

How it works

Under the hood: ffmpeg extracts the audio to a small mono 16kHz WAV, Gemini transcribes with precise timestamps, and the post-processing pipeline runs in order. If you chose burn_in_only or both, it re-encodes the video with the subtitles overlaid - if an ASS file was produced it uses that for full styling, otherwise it applies force_style to the SRT. Files land in PROJEKTS/project/assets/auto_srt/ following the pack's VFX folder convention.

Setup

Same pack install as everything else here - ComfyUI Manager, search comfyui-digit, restart. Two environment prerequisites:

# Gemini needs GCP Vertex auth
gcloud auth application-default login
# ffmpeg must be on PATH or nothing works
ffmpeg -version

The GCP setup is the standard DIGIT one: enable aiplatform.googleapis.com, set gcp_project_id on the node (or rely on auto-detect), leave gcp_region at global. If you see a "File not found" that's the ffmpeg check or the path being wrong - and if audio extraction fails, ffmpeg isn't finding the codec, which is a system-level problem, not a node bug.