| Get the models first | 3 |
| What it does | 4 |
| How many nodes? | 5 |
| Quick start | 6 |
| The MTV node | 7 |
| Casting and Direction | 8 |
| The Cutting Room | 9 |
| Post-Production | 10 |
| Undercover | 11 |
| Under the hood | 12 to 13 |
| Tested, and made with Claude | 14 |
| The man behind it | 15 |
| Troubleshooting | 16 |
| Dependencies | 17 |
Turn your song into a music video, one clip at a time, without building a 50-node workflow.
The main diffusion model is unique to this node pak. Easiest start: open ComfyUI's own MiniMax H3 Ref2V template and the LTX 2.5 template once; they download the official files for you. The author's repo only holds the custom files that are not in those templates.
| File | Folder | Where |
|---|---|---|
| IronLeo78_R2V4Vendetta_ (main model) | diffusion_models | Ironleo78/IronLeo78-MTV |
| h3_upscaler_lms_v0.1 | latent_upscale_models | IronLeo78-MTV / Latent_Upscaler |
| Huihui-Qwen3-VL-4B (Casting LLM) | text_encoders | IronLeo78-MTV or ahmed22xa |
| minimax_h3_lynnreal_light_vae | vae | Kijai / IronLeo78-MTV (backup) |
| qwen3vl_32b_minimax_h3 (text encoder) | text_encoders | Comfy-Org/MiniMax-H3 |
| minimax_h3_audio_vae_fp32 | vae | Comfy-Org/MiniMax-H3 |
| Turbo 4-step LoRA (ref2v) | loras | Comfy-Org/MiniMax-H3 |
| yue2_3b_int8_convrot, sheetsage2_bf16 | checkpoints, audio_encoders | Comfy-Org/YuE2 (Undercover) |
huggingface.co/Ironleo78/IronLeo78-MTV. LTX 2.5 files: huggingface.co/Lightricks/LTX-2.5 (gated: request access).
Pick a song, press Run. The node cuts the matching slice of your audio and generates the picture to follow the music. What you hear is your recording.
One clip per Run. A segment panel keeps your place; every song has its own counter. Go to any segment or time, redo a bad clip.
Continuity. Keep Rolling starts each clip from the last frame of the one before.
Second Unit Director extends the clip about 20 seconds with LTX 2.5.
Sliding modules. Casting and Direction on the left, the Cutting Room on the right, Post-Production docked beside it.
Four themes, saved prompts, 73 cover styles, 1,270+ checks.
You see 3 nodes and work with 1.
| Node | What it is |
|---|---|
| IronLeo78 MTV | The one you work with: song, quality, buttons, segment panel. |
| IronLeo78-Post-Production | Refines and upscales a clip. Spawned beside the MTV. |
| IronLeo78 Undercover | Stand-alone song-cover maker. |
Six node classes are registered. Casting and Direction and the Cutting Room are real nodes too, kept hidden and wired by the MTV, and shown as slide-out panels.
The MTV alone stands in for 30+ stock nodes. Post-Production replaces 20+, Undercover 10+. About 60 nodes and all their wiring become the ones you touch.
1. Put the folder ironleo78_mtv in ComfyUI/custom_nodes/.
2. Get the models (page 3). Restart ComfyUI, press Ctrl+F5.
3. Add IronLeo78 MTV (IronLeo78 > MTV).
4. Custom Audio File: pick your song.
5. Press Casting and Direction: load your singer in slot 1, write a short prompt.
6. Pick Test (fast) or Prod, keep Rolling!, press Run.
7. The clip plays in the Cutting Room. Like it? Switch Remastering on and Run Post-Production.
| Control | What it does |
|---|---|
| Test / Prod | 0.2 MP, 4 steps / 0.4 MP, 6 steps. |
| Custom Audio File | Your song. Its name names the output folder. |
| Choose Audio latent | Custom Audio, MiniMax Audio, or Audio From Video. |
| duration | Seconds per clip (default 5). |
| Keep Rolling | Rolling! next segment every Run. Cut! same segment again. |
| Second Unit Director | LTX 2.5 extends the clip about 20 s. |
| Remastering / Rough Cut | Mirror of the Post-Production switch. |
| Seed Random / Fixed | New seed per clip, or the same one. |
| Segment panel | Go, Go time (1:23), Latest, Reset. |
Outputs: video (to Save Video), filename, status.
Alone: a song and a picture are enough. Together: the hidden modules are already wired to it.
Slides out of the left side of the MTV.
Pictures: a 2 x 2 grid. Click a cell to pick or upload, drop files, red x to remove.
Mode: v2v swaps things in a video you load; i2v starts from pictures.
Preset: plain .md files in presets/. Edit them or add your own.
Prompt enhancer: Qwentin Tarantino's Script = the LLM writes the prompt. Your Script = your text goes through as typed.
Saved prompts: a list above the prompt box. Save prompt as... keeps one as a .txt in prompt_presets/.
Themes: four big swatches at the bottom: Rose Delight, 80s, Corporate, Pioneer.
Slides out of the right side. Nothing to wire.
Video Preview: plays every clip with its sound. Runs list (newest first), frame stepping with , and . keys, a Seam check (2 seconds each side of a join), Save, Delete, rating 1 to 10.
Previews: the live sampler picture with step, speed, time left and a replayable filmstrip.
Refines and upscales one saved clip, x2.00 on each side. Docks to the right of the Cutting Room and follows the MTV.
1. Make clips with the MTV (latents are always saved).
2. Pick the latent (newest first; press R). The conditioning is found by itself. Upload brings clips from another machine.
3. Switch Remastering on and Run. Link video to Save Video. It then flips back to Rough Cut, so a new scene never starts a refine.
4. Delete all latents clears every clip file. Press twice. Videos stay.
It refines with the clip's own seed, with no LoRa, at euler and 3 steps.
Always stand-alone. Load a song, write your lyrics, pick a style: you get a YuE2 cover saved as FLAC or MP3 320k in ComfyUI's input folder.
SheetSage2 reads the melody. Key and BPM can change: the tune is transposed, not replaced.
73 style presets (ported from SongScribe by TheLocalLab, MIT) with Vocals, Era, Texture, Mood Shift, Language and Blend.
Analyse only reads key, BPM and meter without making a cover.
One engine at a time. The MTV and Post-Production share one cached H3 base model. The Casting LLM, LTX and YuE2 give memory back before H3 loads.
The MTV Run:
1 wires resolved, 2 saved seed applied, 3 models load (cached), 4 conditioning made with ComfyUI's own MiniMax node and cached, 5 sample (BasicGuider, res_multistep), 6 decode and create the video (24 fps, h264), 7 latents, sidecar and manifest saved, 8 Seed Random picks the next seed, the counter moves on.
Segments: segment 1 starts at second 0; each next one a real clip length later (frames / 24). A clip is frames % 17 == 5 long: 5 s = 124 frames.
Why audio is frozen: your song's latent gets a zero noise mask, so sampling never touches it. The picture follows your recording.
Files a clip leaves: .mp4, _video.latent, _audio.latent, .conditioning, .json sidecar, plus the song's manifest CSV. The latent header holds the seed.
Post-Production: check, load, upscale x2.00 (trained node if installed, else built-in resize), build model, 3 euler steps (0.9035, 0.6316, 0.3158, 0), decode, create video. Rough Cut returns blockers and does nothing.
The UI: plain ES modules. Slide-out panels are children of the node's own element. Hidden ghost nodes carry the real widgets. Values are restored by name. Colours come from one contrast-tested theme table.
Hardcoded on purpose: diffusion model, VAEs, text encoder, attention, upscaler, refine schedule, Casting LLM, LoRa (off), 24 fps. Fewer knobs, fewer bad clips.
27 test files, 1,270+ checks: Python against stand-ins for ComfyUI (no GPU), JavaScript in jsdom, a preset audit, a builder check, ruff, and a syntax check of every web file. Every release must be ALL GREEN before it is committed, archived and installed.
A reviewer (Gemini) once called the code "industrial grade solid" and robust. Take that lightly, lol. The tests back it up, and so would Claude.
Built with Claude Code, and written to be changed with it: ARCHITECTURE.md is the map, CHANGELOG.md lists every release, and every release ships handoff notes so a fresh session knows where things stand.
Want another button name, theme or preset? Open the folder in Claude Code and ask.
IronLeo78 spent 25 years in metal music, and has digital video editing experience on top: cutting picture to sound is the whole point of this node.
A deep passion for music, and for directors like Terry Gilliam and Quentin Tarantino (to name a few), is why the UI talks like a film set: Casting and Direction, the Cutting Room, Post-Production, Keep Rolling, Second Unit Director, Remastering, Rough Cut.
More than 55 releases, about a hundred commits, many iterations and a lot of "no, not like that". The design you see is what survived.
| Problem | What to do |
|---|---|
| Nodes missing | Folder directly in custom_nodes with __init__.py. Restart, Ctrl+F5, check the console. |
| Import error (MiniMax H3) | Update ComfyUI. |
| Model not found | Exact file name in the right models folder, press R. |
| Panels or buttons look wrong | Use Nodes 2.0, Ctrl+F5. |
| Segments: unknown | Pick a Custom Audio File or Video File. |
| Node stops and does nothing | Last segment passed: Go, Latest or Reset. |
| Out of memory | Close GPU programs, use Test, retry. |
| Post-Production list empty | Make a clip, press R, or Upload. |
| "audio latent missing" | The three clip files must share a name and folder. |
| Post-Production does nothing | Remastering is OFF, or it already ran once. |
| Phone video on its side | Turned upright by itself; mirrored video is not un-mirrored. |
ComfyUI recent enough to include the MiniMax H3 nodes, LTXVSeparateAVLatent / LTXVConcatAVLatent, CreateVideo, SaveVideo, ModelAttentionBackend (optional) and the core samplers and loaders.
Python packages: none to install. torch, torchaudio, numpy, Pillow, PyAV, safetensors and aiohttp ship with ComfyUI. No ffmpeg.
Optional: the Comfyui_Minimax_h3_latent_Upscaler node (MIT, LBH-123-AI) for the trained upscaler:
Hardware: an Nvidia GPU (the upscaler runs on CUDA). H3 and LTX are never in RAM together: about 79 GB peak with Second Unit Director.
Gated: LTX-2.5 needs access on Hugging Face. Qwen3-VL 4B / 8B encoders do not work as the H3 text encoder.