ComfyUI_LC_AV_nodes

ComfyUI LC Audio_Video Nodes

Custom nodes for ComfyUI by lonecatone23.

• Repo: https://github.com/lonecatone23/ComfyUI_LC_AV_nodes
• Image companion: https://github.com/lonecatone23/ComfyUI_LC123_nodes
• CivitAI: lonecatone23
• Support: https://ko-fi.com/lonecatone
• Version: 0.5.20 · 27 nodes

Small tools that remove friction: frames, soundtrack, metadata, fewer extra packs.

This pack is not a fork of LC123. Do not merge the two folders. LC_PIPE stays on the image pack. This pack uses LC_AV_PIPE.

Release history lives in git tags. This page describes the pack as it is now.

---

Install

ComfyUI/custom_nodes/ComfyUI_LC_AV_nodes/__init__.py

__init__.py must sit directly in that folder, not in ComfyUI_LC_AV_nodes/ComfyUI_LC_AV_nodes/. If you unzip a pack zip inside an existing clone, move the inner files up one level.

Restart ComfyUI. Console should print [LC AV] total … nodes. Hard-refresh the browser after a web/ JS update.

Requirements: ComfyUI’s Python (torch, numpy). Save Video needs ffmpeg on PATH. Audio Separate needs torchaudio (no Numba / SoX).

Optional example graph: workflows/LC_AV Node Examples.json
(BETA: examples only. Add nodes from the node menu; do not treat the graph as a production template.)

---

What this pack is

| Title | Class | Role | | --- | --- | --- | | LC AV Pipe (in/edit) 🎬 | LCAvPipeEdit | Pack or edit LC_AV_PIPE. In: pipe, av_pipe, then video / audio / fps / … Out: av_pipe. | | LC AV Pipe Out 🎬 | LCAvPipeOut | Unpack. In: pipe + av_pipe. No LC_PIPE on the output side. | | LC Pipe → AV Pipe 🔌 | LCPipeToAvPipe | One LC_PIPE in, one LC_AV_PIPE out. Shared keys only. | | LC Create Video 🎬 | LCCreateVideo | IMAGE batch + fps (+ optional AUDIO) → LC_AV_PIPE. Does not encode a file. | | LC Save Video 💾 | LCSaveVideo | ffmpeg encode. Inputs: images, audio, av_pipe, metadata. No LC_PIPE socket. | | LC Load Video 🎞️ | LCLoadVideo | A video file in, its frames, sound and frame rate out, exactly as the file has them. Streams with a progress bar; Cancel works. | | LC Preview Video 📺 | LCPreviewVideo | Plays a video file, or frames with their sound, on the node. Nothing saved to output. | | LC Save Video Metadata 🏷️ | LCSaveVideoMetadata | LC_SAVE_META + A1111 parameters. Works with Save Video or LC123 Save Image. | | LC Final Frame 🖼️ | LCFinalFrame | Last frame of an IMAGE batch. | | LC VAE Decode 🧩 | LCVaeDecode | Boolean wrapper: core VAEDecode or core VAEDecodeTiled. | | LC Get Latent Size 📐 | LCGetLatentSize | Pixel width / height, batch, frames. Optional VAE for compression. | | LC Load Audio 📂 | LCLoadAudio | Load from input. start_time / end_time cut, preview / pause, timeline handles. | | LC Audio Duration ⏱️ | LCAudioDuration | Seconds int + float (and minutes). | | LC Audio Crop ✂️ | LCAudioCrop | Crop AUDIO with MM:SS start / end. | | LC Audio Channel 🎚️ | LCAudioChannel | Stereo → L / R. | | LC Audio Separate 🎼 | LCAudioSeparate | Demucs stems: Bass, Drums, Other, Vocals. | | LC Audio EQ 3-Band 🎛️ | LCAudioEqualizer3Band | Low shelf, mid peak, high shelf. | | LC Audio Volume 🔊 | LCAudioVolume | Gain 0–4, soft clip. |

---

LC_AV_PIPE 🎬

Type string: LC_AV_PIPE.

Edit node sockets, top → bottom:

• pipe (LC_PIPE)
• av_pipe (LC_AV_PIPE)
• video, audio, frame_rate, duration, sample_rate, width, height, frame_count
• positive / negative, seed, total_steps, sampler_name, models

Wired slots override; empty slots pass through.

Do not put video into LC_PIPE. Convert shared fields with LC Pipe → AV Pipe.

No civitai_air on the pipe. Paste AIR on Save Video Metadata.

---

Load Video 🎞️

Path in (a full path, or one inside ComfyUI's output or input folder): every frame at the file's own size, its sound and its frame rate out. Nothing to set: no resizing, frame skipping or trimming. Frames stream in through ffmpeg with a progress bar, and Cancel stops it. Wire LC H3 Make's video straight in.

Preview Video 📺

Plays a video on the node: a file (path), or a batch of frames with its sound (fps sets the rate). Frames are encoded to a temp file, so nothing lands in your output folder.

---

Save Video 💾

VHS-style options: frame_rate, loop_count, filename_prefix, format, pix_fmt, crf, pingpong, save_output, prune_outputs.

prune_outputs keeps intermediate frame PNGs off disk.

save_metadata writes sidecar JSON + container comment. CivitAI reads PNG chunks, not the mp4 comment. Keep civitai_still on and upload the first-frame PNG (fill models + civitai_air on the metadata node).

Frames stream into ffmpeg one at a time with a progress bar, so long clips don't fill your RAM, and Cancel stops it (no half-written file is left).

On-node preview after encode. Resize the node; the player follows aspect.

Wire IMAGE from VAE Decode (or Create Video images) into images. Metadata is optional. av_pipe can carry frames if they were packed there.

---

LC VAE Decode 🧩

tiled off = nodes.VAEDecode.
tiled on = nodes.VAEDecodeTiled (tile_size, overlap, temporal_size, temporal_overlap).

Same latent + same VAE as the two core nodes. Use the MiniMax Video VAE for video latents.

Shows a progress bar per tile (also on the MiniMax Video VAE, which tiles itself), and Cancel stops it between tiles.

---

Load Audio 📂

Load audio uploads into Comfy input/ like stock Load Audio.

start_time / end_time (seconds). end_time 0 = end of file. Output duration is the cut length.

preview / pause plays only that cut. Drag the two handles on the timeline (after one preview so file length is known).

---

H3 Long Video 🎬

Minute-long MiniMax H3 videos made in short pieces, one shot at a time, on the same card that makes 5 s clips. LC H3 Story 🎬 holds the cast, places and shots. LC H3 Make 🎥 renders them and joins the movie. Example workflow: MMH3 Showrunner V2.5.2 (with a guided tour: Vision Assistant → Tour).

📘 User Guide: every part of the workflow, step by step, with screenshots. ✍️ Prompting Guide: how to word looks, outfits, actions and lines.

LC H3 Story (BETA) 🎬

• The whole movie on one node: cast, outfits, places, props and shots, each on its own card.
• ✋ The text a shot builds is exactly what H3 gets. Open the shot's H3 prompt to read it. One thing is left out quietly: a place's sound line loses any clause about speech ("nobody speaks"), because naming speech makes H3 speak. You see the result in the H3 prompt.
• Cast card: a name, who they are (woman, man, cat, etc.), their look, a headshot (head and hair only), a 👗 outfit picture and a voice.
  • 💡 Use a neutral face on a plain grey background. A card's background, props and expression leak into every shot it's in.
• Outfit card: a change of clothes, with a simple name ("red cocktail dress"), who wears it, a head-to-toe description and a picture. A state of dress works too ("naked", "topless, black jeans", "wrapped in a white towel"): H3 gets it as written.
• Prop card: an object that matters (a cup, a sign), with its own picture.
• 📚 Every cast, outfit, prop and 🌐 360 slot has a Library tab: 102 ready-made pictures, free (CC BY 4.0).
• Place card: a description, the sound of the room ("rain on the window, cups clinking, footsteps": sounds only, never who speaks) and a picture of it empty. A 360 panorama lets close-ups face any side of the room: aim them with the 🌐 viewer on the shot.
• Shot card: framing, who it's on (everyone, one person, or nobody for an empty shot), camera move, cut or keep rolling, the action, who speaks and the line.
  • 🎞 keep rolling carries straight on from the shot before and locks its framing, place, 🌐 angle, outfits and start frame (🔒). People can still walk in, and they get their face pictures.
  • 🎛 LoRA for this shot: any LoRA for that shot only, on top of the pipe's LoRAs, picked like the LC LoRA Loader (folder picker, strength, ℹ trigger words). Keep rolling uses its cut's LoRAs.
  • seconds: auto fits the beats and the line, or type your own. The number you type wins.
  • Name (V.O.) in the speaker list is a voiceover: heard off screen, lips stay closed.
  • 💡 One line over a cut: end it with ... on this shot and start the next shot's line with .... The next shot picks the line up straight away.
  • 🖼 start / 🏁 end pin the first or last frame. H3 copies a pinned frame exactly, so make it sharp.
  • 📐 draws boxes: who stands where. ✨ on the 🖼 start slot turns them into a start frame.
  • Count props once ("each holds one cup") or extras show up.
• ✨ Write the story turns a short pitch into cards with a local LLM, once. After that it's plain text you edit.
  • 🔞 NSFW (off by default): lets the writer write adult scenes, the same rule as LC Vision's NSFW style tag.
• The board under the shots: a tile per shot. ✅ made, ✏️ changed since it was made, ⚠ join broken (a keep-rolling shot whose shot before was remade), 🔒 locked. ▶ plays it, 🩹 repairs it (a new take that keeps its last moment, so the shot after it still joins), ↻ makes that one shot again and stops, 📌 on a paused frame sets that place's room anchor, 🖼 on a paused frame starts the next shot from it.
• ▾ folds a shot to one line. The project bar stays pinned at the top.
• max_shot_seconds: the longest shot this machine renders in one go (20 default). A longer shot is made in parts, each carrying on from the one before.
• H3 prompt (under each shot): the shot's lines (Place, Camera, the timed action). The people, picture and sound sections are added around them when the shot is made. ↻ builds it from the cards again right now. ✎ Edit lets you write the lines by hand: they go in as written, and changes to the cards stop changing it until you press ↶ Back to automatic.
• ✨ on a picture slot (cast, outfit, place, 🌐 360, prop, 🖼 start, 🏁 end) makes that picture. The Story switches on the part of the workflow that makes it, keeps Make off, runs once and puts everything back. A blank card name never writes over a picture, and ↶ brings back the picture before (the last 8 are kept).
• 📐 size (under the project name) shows the size the project is made at. Change it and the next Run makes every shot made at the old size again. 💾 Save as keeps a variant under a new name. 🆕 New clears the board (it warns first when there is unsaved progress).
• The 🌐 shot angle is also the background of the 📐 boxes, so you place people on the view the shot will use.
• 🌐 360 from card (under a place's room picture): LC Vision reads the room picture and writes the whole place out, then ✨ draws the 360 from that. Needs LC Vision.
• Everything saves to output/LC_H3/<project>/ (story.json, plan.json, the takes and the movie). Pictures go in input/lc_h3_refs/<project>/.

The picture makers (inside the Cards & Storyboard 🃏 subgraph of the example workflow; you run them with ✨ on the Story)

• LC H3 Story Card Prompt: a card's description as an image prompt and an empty latent at the size that picture needs (headshot, outfit, room, 360 panorama, prop). kind and name are set by ✨.
• LC H3 Shot Card: one shot's start or end frame as a storyboard job for any image model: the prompt, the 📐 boxes, the background (the shot's 360 view or the room) and the cast pictures.
• LC H3 Save to Story: puts the finished picture on its card (from card), or on a shot's start or end frame. A new take gets a number, so older pictures are kept (↶).
• Any image model works: the model sits in the workflow (the LC Model Optimizer and the LoRA loader in front of the subgraph), never in these nodes.

LC H3 Make (BETA) 🎥

• Renders the Story's shots one after another and joins them into the movie. Wire in the Story's plan and refs, and an LC MiniMax H3 Pipe V2 (model, text encoder, both VAEs, size, steps, sampler, seed).
• make: all remaining (default) makes the whole movie in one Run. next segment makes one. redo makes segment number index again, with a new seed.
  • 💡 One Run does it all. Stop ends it with nothing left in the queue. Run again to carry on.
  • Edited a shot that's already made? Run remakes only the ones marked ✏️, the rest stay.
• redo_after: redoing a shot also sets aside the keep-rolling shots after it, since they no longer join. The old take is always kept in takes/.
• refs_on: cuts only (default) gives the smoothest joins. A keep-rolling shot carries on from the frames before it, and a fresh reference pulls it away.
• hold_scene (on): the first wide shot of a place becomes its layout, so later shots there keep who sits where.
• memory: clip (default) feeds about 1 s of the first shot back on every cut so faces, clothes and voices stay put. memory_gap 5 s stops the cut from opening on that clip's pose.
• color_match: 0 = off. 0.2 to 0.4 when info flags a scene drifting warmer or cooler.
• crf 16 is close to lossless.
• info reports the face match per shot (0.55 and up = same person, under 0.40 = someone else, redo it) and how smooth each join is.
• A shot's 🎛 LoRAs go on a copy of the model for that shot only; the next cut starts from the pipe's model again.
• 💡 960 x 544 is the low VRAM size. Frame Interpolate to 48 fps is the post step that clearly helps; upscalers were barely worth it in my tests.

LC H3 Audio Sync (BETA) 🎤

• Makes an H3 video perform YOUR track (a song, a line of dialogue) instead of sound the model invents. Put it between the latent and the sampler.
• Your recording is locked into the sound half, so the model only decides the picture that fits it, mouth included.
• 💡 Wire the H3 positive conditioning through it too: the track is also handed to H3 as a clean guide.
• ✋ Put each line in the prompt with its time: At 0 seconds she says, "Hello there." H3 moves the mouth when the prompt says the words are spoken, not when it hears them. Without the time she talks over the silence instead.
• start_seconds picks where in the track the clip starts. A track shorter than the clip is filled with silence or looped.
• The audio output is the track cut to the clip's length: wire it to the save node so picture and sound end together.
• ⚠️ H3 was trained to about 15 s; longer clips get a warning in info.

LC Frame Interpolate 🎞️

• Smooth motion with RIFE 4.26: makes the in-between frames so a 24 fps clip plays at 48, 60, or any rate (not just 2x). Each new frame lands at its real moment in time.
• scene_cut holds hard cuts instead of morphing across them (a morph across a cut is a smeared ghost frame). 0 = off.
• Wire the clip's audio and the output is exactly as long as the sound, so picture and audio end together.
• model: rife426.pth (best on AI video in our tests, downloads once, 21 MB) or crossfade (no model, fast, soft on motion). Uses models/rife/ or ComfyUI-Frame-Interpolation's checkpoints if you have them.
• RIFE code from ComfyUI-Frame-Interpolation (Fannovel16) and Practical-RIFE (hzwer), both MIT.

💡 Vision Assistant → Movie (LC Vision) shows all of it in plain words, with one-click redo and Make all remaining.

---

Colors

web/lc_node_colors.js: AV family teal (#2a6a6a audio, #1a5a7a video). Distinct from LC123 sigma #1c6d6d. VAE Decode uses stock Comfy red.

web/lc_color.js is the launch-color helper (does not overwrite a saved color).

---

License

MIT. See LICENSE.