🎞 Save Video (grade) - LD

The last node in a video workflow is usually a boring Save Video: frames in, mp4 out. Save Video (grade) - LD is the opinionated version. It's the pack's finishing node - category LD / PromptMaster / Finishing - and it bundles the two things that separate a render you'd post from a render you'd delete: a grade you can preview instantly, and an encoder step that's been tuned against real measurements rather than vibes. It's an output node, so it sits at the end of the rail.

What it actually does

Executing the node stages a temporary preview in its own panel - instant, because the preview is CSS filters over the frames rather than an encode. Pressing SAVE is what writes a real file to ComfyUI/output/, and that save goes through ffmpeg (system ffmpeg if you have it, else the one bundled with ComfyUI), which is where the actual pixels get decided. save_master (on by default) writes the master to output/ on every run, so this node can be your only save node and nothing is lost when a tab closes.

It takes two render rails because the pack runs two different samplers: images + audio for single-shot video (T2V / I2V / REF), and images_b + audio_b for chained H3 Master output (MULTISHOT / MV). Lazy evaluation only pulls whichever rail is actually live, so a dead sampler never gets flattened. Output is a single path string pointing at what you saved.

The controls that earn their keep

• sharpen (0–1, default 0) - contrast-adaptive sharpening. VAE decode is soft; 0.3–0.5 is plenty, and CAS leaves contrasty edges alone so you don't get white halos around hairlines.
• grain (0–60, default 0) - film grain, applied after sharpening so it's never sharpened into crawling noise. 4–10 is subtle, 20+ is a look. This is the classic anti-"AI look" move, and the KB's own corpus shows it's folk-standard practice for a reason.
• crf (default 19) - x264 quality, where lower is cleaner and bigger. The tooltip carries the author's actual measurements, and they're the most useful thing on this node: grain is where size goes, because random noise is incompressible - at crf 17, a grain-8 clip cost 98× the no-grain size; at crf 20 it was 13×. So 15–17 for clean plates with no grain, 19–20 for normal, 22+ when size matters more than the last of the detail.
• encoder - h264 (plays anywhere), h265 (same look, ~30% smaller), or av1 (smallest, and the only 10-bit option a browser actually plays; slower to encode). bit_depth stays at 8, and that's a decision with a reason: 10-bit H.264 measures better on paper but doesn't play in Discord, browsers or the Windows player, and a master nobody can open isn't a master.
• interp_mult - this one trips people up. It does not interpolate; if a RIFE node upstream doubled your frames, you set this to the multiplier so the file is written at the doubled fps - smoother instead of 12 seconds of slow-motion with the audio adrift.
• scale - a Lanczos resample of the saved master. It's a resize, not an upscaler (no invented detail), but it can't flicker the way a per-frame neural upscaler can. For real added detail the tooltip's advice is blunt and right: run SeedVR2 upstream and leave this at 1.
• loudnorm - EBU R128 to −16 LUFS, the streaming target. H3's re-synthesis and a muxed original MP3 land at very different levels, so without it a run of takes jumps in volume.
• mux_mp3 - filename in ComfyUI/input of the original track. When set, the saved file carries that audio instead of H3's re-synthesis, at offset zero, because the picture was generated against that exact waveform.

Install and the honest catches

cd ComfyUI/custom_nodes
git clone https://github.com/Lora-Daddy/PromptMasterLD
# restart ComfyUI

It needs an ffmpeg binary somewhere reachable - ComfyUI ships one via imageio, so usually you're fine; if you see "no ffmpeg found", install one on PATH or pip install imageio-ffmpeg into ComfyUI's python. The one real footgun: mux_mp3 only lands right when the chain's frame maths is correct - a clip that ships fewer frames than its window skips the song at every join, and no fixed offset fixes that. Keep save_master on, read the preview as an approximation (CSS has no gamma or color temperature, so close, not exact), and trust SAVE for the real thing.