IronLeo78 MTV
Turn your song into a music video, one clip at a time: the IronLeo78 MTV node (MiniMax H3) with its Casting and Direction and Cutting Room modules, Post-Production (refine and upscale). Use with Nodes 2.0.
Nodes (10)
IronLeo78 MTV ((v1.5.0))
USE NODES 2.0!
Using the original (classic) nodes is discouraged: this pack must be used with Nodes 2.0. It is WORTH it: the pack pushes the beast features of Nodes 2.0 further than a rogue programmer usually dares. The slide-out panels, the docked Post-Production, the theme tint and the buttons only work there. When an IronLeo78 MTV node opens and Nodes 2.0 is off, the pack switches it on for you and shows a notice (if the node looks odd, press Ctrl+F5). To keep the classic canvas, run
localStorage.setItem("ironleo78_keep_classic", "1")in the browser console: you then only get the notice.
Get the models first. The main diffusion model is unique to this node pack. Exact file names are fixed in the code. Folder = under ComfyUI's
models/.Easiest way to get the basic weights: open ComfyUI's own MiniMax H3 Ref2V template and the LTX 2.5 template once. They download the official files for you automatically. Starting there is the recommended way. The official LTX and MiniMax weights are not hosted in the author's repo. The author's repo https://huggingface.co/Ironleo78/IronLeo78-MTV only holds the custom files that are not in those templates (the main diffusion model, the latent upscaler and the Casting LLM).
| Model (file) | Folder | Where | |---|---|---| |
IronLeo78_R2V4Vendetta_.safetensors(main diffusion model, the author's own) |diffusion_models| file / repo | |minimax_h3_lynnreal_light_vae_int8_convrot.safetensors(video VAE) |vae| Kijai/MiniMax-H3-experimental/ backup copy | |qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors(text encoder) |text_encoders| Comfy-Org/MiniMax-H3 | |minimax_h3_audio_vae_fp32.safetensors(audio VAE) |vae| Comfy-Org/MiniMax-H3 | |minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors(Turbo LoRA) |loras| Comfy-Org/MiniMax-H3 | |h3_upscaler_lms_v0.1.safetensors(latent upscaler, used by Post-Production) |latent_upscale_models| IronLeo78-MTV | |Huihui-Qwen3-VL-4B-Instruct-abliterated-int8_convrot.safetensors(Casting LLM) |text_encoders| IronLeo78-MTV or ahmed22xa | |taeh3.safetensors,taeltx2_3.safetensors(optional sharper preview) |vae_approx| madebyollin/taehv | | LTX 2.5 files for the LTV node (Gemma encoder, video and audio VAE; gated, request access) | seeltv.py| Lightricks/LTX-2.5 |If a model is updated or removed upstream: the node only looks for the exact file names in the table. A newer version of a model works if you rename the file to the name the node expects (keep the original name). The author keeps a backup copy of the Lynn video VAE (
minimax_h3_lynnreal_light_vae_int8_convrot.safetensors) in case Kijai's upload changes or disappears.
Turn your song into a music video, one clip at a time, without building a 50-node workflow.
Release candidate territory: an everyday-tested ComfyUI node pack for MiniMax H3. A singer or musician performs your track, and every clip is cut to your own recording.
Contents
- Summary of features
- How many nodes?
- The work behind it
- Tested
- Made to be changed with Claude
- The man behind it
- Quick start
- The nodes, one by one
- Under the hood
- Troubleshooting and dependencies
Summary of features
- Pick a song, press Run. The node cuts the matching slice of your audio, freezes it into the video latent and generates the picture to follow the music. MiniMax H3 never invents sound: what you hear is your recording.
- One clip per Run, and it keeps its place. A segment panel shows how many segments your song holds, which is next, and your progress. Jump to any segment (or to a time), redo a bad one, or let it step through the whole song. Every song keeps its own counter.
- Continuity built in. With Keep Rolling ON, each clip starts from the last frame of the previous one, so the performance flows.
- Second Unit Director: after the H3 clip, LTX 2.5 can extend it by about 20 seconds in the same Run.
- One small node, hidden helpers. The Casting and Direction panel (pictures, mode, preset, prompt, saved prompts, themes) slides out of the left side of the node. The Cutting Room (live preview and clip player) slides out of the right side. Both are always active, just out of sight until you want them.
- Post-Production (refine and upscale x2) docks to the right of the Cutting Room and follows the node when you drag it, like one big node.
- Casting LLM: a Qwen3-VL 4B writes the MiniMax prompt from your pictures, your video and a short request. Every preset is a plain
.mdfile you can edit. - Film-set buttons: Keep Rolling (Rolling! / Cut!), Test / Prod, Second Unit Director, Post-Production / Back To Prod, Seed Random / Seed Fixed, Qwentin Tarantino's Script / Your Script.
- Saved prompts: keep the prompts that worked as plain
.txtfiles and pick them from a list. - Four themes: Rose Delight (default), 80s, Corporate, Pioneer.
- Everything is saved: each clip leaves its video, latent, audio latent, conditioning and a
.jsonsidecar with the full prompt and settings, plus a manifest and a run log. - Tested on an RTX 5080 16 GB, plus more than 1,270 automated checks.
How many nodes?
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 | the Revitalizer: refine and upscale a clip (spawned beside the MTV) |
Inside the pack, six node classes are registered. The Casting and Direction and the Cutting Room (Video Preview + Previews) are real nodes too, but the MTV keeps them hidden and wired, and shows their controls in its slide-out panels.
Counted from what the code calls, the MTV alone stands in for 30+ stock nodes (model, LoRA and attention loaders, text encoder and VAE loaders, audio / video / picture loaders, MiniMax H3 reference-to-video and guide, audio trim and encode, masks, AV concat / separate, noise, guider, sampler select, scheduler, sampler, decoders, Create Video, latent savers). Post-Production replaces 20+ more. About 60 nodes and all their wiring become the ones you actually touch.
The work behind it
This pack grew one small request at a time, every day, through more than 55 numbered releases and about a hundred commits. Every release is tested, committed, zipped with its handoff notes, and installed. Features were tried, kept, thrown away and redone: a merged all-in-one node was built and thrown out, the Model Loader was folded away, the Revitalizer's sigma buttons came and went, presets were rewritten many times. The design you see (one node, two sliding modules, one docked helper) is the result of a lot of "no, not like that".
The fine tuning never stops: the quality settings were compared blind, the refine pass was frozen to the settings that looked best, and the buttons got their film-set names because the work should be fun.
Tested
python tools/run_all.py # everything, about 5 minutes
python tools/run_all.py --quick # seconds
- 27 test files, more than 1,270 checks: Python tests against stand-ins for ComfyUI, JavaScript tests in jsdom (the buttons, panels, themes, sync logic), 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.
- It is also tested the old way: the author uses it daily in a live ComfyUI, on a real GPU, with real songs.
An AI reviewer (Gemini) once called the code "industrial grade solid" and robust. Take that lightly, lol. But the test suite does back it up (and so would Claude).
Made to be changed with Claude
The whole pack was built with Claude Code, and it is written so Claude can keep changing it:
ARCHITECTURE.mdis the map: which file owns what, the traps that cost time once, what is hardcoded on purpose.CHANGELOG.mdlists every release.- Every release comes with handoff notes (
HANDOFF_x_y_z.md, inside the release zip too) so a fresh Claude session knows where things stand. - The tests protect the design: ask Claude for a change, it runs
tools/run_all.py, and anything it breaks is caught. - Presets, guides and saved prompts are plain text files.
Want a different button name, another theme, another preset? Open the folder in Claude Code and ask.
The man behind it
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, Post-Production, Back To Prod.
The pack is MIT licensed and shared in that spirit.
Quick start
- Install the pack: put the folder
ironleo78_mtv(with__init__.pydirectly inside) inComfyUI/custom_nodes/. - Get the models (table at the very top). Easiest: open ComfyUI's own MiniMax H3 Ref2V template and LTX 2.5 template once, then add the custom files from the author's repo.
- Restart ComfyUI and hard-refresh the browser (Ctrl+F5). Use Nodes 2.0 (the pack switches it on by itself when the MTV opens).
- Right-click the canvas, Add Node > IronLeo78 > MTV > IronLeo78 MTV.
- Custom Audio File: pick your song.
- Press the Casting and Direction button at the top of the node to slide that panel out: load your singer's picture in slot 1, write a short prompt (for example The singer sings to the camera, stage lights).
- Pick Test (fast, small) or Prod, leave Keep Rolling on Rolling!, and press Run. One clip per Run. The clip appears in the Cutting Room panel on the right.
- Like a clip? Pick it in Post-Production and press the big SELECT YOUR LATENT AND CLICK TO REMASTER button: only that node runs. Or click Post-Production on the MTV and Run: its panel opens and every other node is bypassed, so it refines whatever is selected in that list and nothing else. It stays in Post-Production until you press Back To Prod on the Post-Production node.
The nodes, one by one
All nodes live under IronLeo78 > MTV. Each one also works alone.
IronLeo78 MTV: the clip maker
One clip per Run, cut and synced to your song.
| Control | What it does |
|---|---|
| Test / Prod | Test = 0.2 MP, 4 steps (fast checks). Prod = 0.4 MP, 6 steps (the real thing). |
| Custom Audio File | Your song. Its name becomes the output folder and file prefix. |
| Choose Audio latent | Custom Audio (default): your song, chunked and frozen into the latent. MiniMax Audio: the model makes the sound. Audio From Video: the soundtrack of the loaded video. |
| Video File | For video-to-video: the one place a video is loaded. |
| duration | Seconds per clip (default 5; real length follows MiniMax's frame rule). |
| Keep Rolling (Rolling! / Cut!) | Rolling! = every Run moves to the next segment and continues from the previous clip's last frame. Cut! = every Run remakes the same segment. |
| Second Unit Director | ON = LTX 2.5 extends the clip by about 20 s in the same Run. Needs the LTX files. |
| Post-Production (panel bar) and Back To Prod (v1.4.0) | Post-Production enters the room: its panel opens, it runs stand-alone on the clip selected in its list, and every other node is bypassed (like Ctrl-B). Back To Prod, on the Post-Production node, goes back to Production: the other nodes come back and the panel closes. |
| Seed Random / Seed Fixed | Random: a new seed after every clip. Fixed (the default): same inputs, same clip. |
| IronLeo78 Tuned / ComfyUI Vanilla (top of the node) | The models of the MTV and Post-Production, one choice for the whole pack. Vanilla (the default): the files of ComfyUI's own r2v template with its 4-step Turbo LoRA at strength 1. Tuned: the author's model and video VAE. Same Test / Prod steps in both. |
| Segment panel | segments in the song, the next one, progress; Go, Go time (type 1:23), Latest, Reset. |
No outputs (v1.3.0): the Cutting Room plays and saves the clip with no wire, and the run report shows in the status box at the bottom of Post-Production. Optional inputs: Pictures (from Casting and Direction), song (any AUDIO), song_name.
Alone: give it a song and a prompt-less picture setup and it still works. Together: the hidden Casting and Cutting Room modules are wired to it already.
Casting and Direction (slides out of the left side)
Pictures, mode, preset and prompt for the MTV.
- 2 x 2 picture grid (click a cell to pick or upload, drop files on it, red x to remove). Empty = NONE.
- Mode v2v / i2v: v2v swaps things in a video you load; i2v starts from pictures.
- Preset: plain
.mdfiles inpresets/(v2v Swap main female actress, Swap main male actor, Change the background). Edit them or add your own. - Prompt enhancer (Qwentin Tarantino's Script / Your Script): ON, the LLM writes the MiniMax prompt from your request, pictures and video. OFF, your text goes through exactly as typed.
- Saved prompts: a list above the prompt box; Save prompt as... keeps the prompt as a
.txtinprompt_presets/(drop your own files there too); Delete removes one. - Themes: four big swatches at the very bottom.
Stand-alone: it is an ordinary node too (IronLeo78_MTV_PromptGenerator): wire its Pictures output to the MTV yourself. Its prompt does not go on a wire (v1.2.0): it reaches the MTV through the pack, and the box under the Preset shows the text Qwen wrote on the last Run.
Cutting Room (slides out of the right side)
- Video Preview: plays every clip with its sound; runs list (newest first, filters), frame stepping (
,and.), Seam check (2 s each side of a join), Save, Delete, rating 1 to 10 (kept in the run log). Clips live inoutput/IronLeo78-Video-Preview. - Previews: the live sampler picture (step, speed, time left, replayable filmstrip). Set ComfyUI's Preview Method to
taesdorlatent2rgb.
Both need nothing wired; they follow the MTV.
IronLeo78-Post-Production (docked right of the Cutting Room)
Refines and upscales one saved clip (x2.00 each side).
- Make clips with the MTV (the latent is always saved).
- In the node, pick the latent (newest first; press R after a new run). The matching
.conditioningis found by itself. Upload latent / conditioning files... brings clips from another machine. - Click Post-Production on the MTV and Run. It stays active so you can refine several clips; press Back To Prod to go back.
- Delete all latents clears every clip file (
.latent,.conditioning) from the output and upload folders (press twice; videos stay).
It refines with the clip's own seed, with no LoRa, at euler and 3 steps. Alone, it works from any saved clip files.
Under the hood
The geek part. (ARCHITECTURE.md has the full map.)
One engine at a time
The MTV and Post-Production share one cached MiniMax H3 base model (model_cache.py). The Casting LLM and LTX give the memory back before H3 loads, and the other way round: only one engine is in VRAM at a time. The LLM can stay in RAM between runs so the next prompt starts fast.
The MTV Run
- The
songandPictureswires are resolved; the Video File decides v2v or i2v. - Hardcoded settings and the saved seed (
mtv_defaults.json) are applied; Test / Prod pick megapixels and steps. - Models load (cached), then media: video frames, the song chunk for this segment, the pictures.
- Conditioning is made with ComfyUI's own MiniMax H3 node (and cached by content), the guide frame is added, the audio latent is frozen with a zero noise mask.
- Sample (
BasicGuider,res_multistep,simplescheduler), decode, create the video (24 fps, h264). - Latents and conditioning are saved next to the clip; the optional LTX extension runs; the manifest, sidecar and status are written.
- Seed Random (if on) saves a fresh seed for the next clip; the segment counter moves on.
The segment cut is deterministic: segment 1 starts at second 0, each next one a real clip length later (frames / 24). A clip is frames % 17 == 5 long (duration 5 s = 124 frames).
Why audio is frozen
With Custom Audio the song's latent gets a zero noise mask, so sampling never touches it: the picture is generated to follow your recording, and the sound in the video is your sound.
Files a clip leaves
output/MySong/MySong_seg004_A_00001_.mp4
output/MySong/MySong_seg004_A_00001_video.latent
output/MySong/MySong_seg004_A_00001_audio.latent
output/MySong/MySong_seg004_A_00001.conditioning
output/MySong/MySong_seg004_A_00001_.json (prompt, settings, file names)
output/MySong/MySong_manifest.csv (start / end second of every clip)
The latent header holds the clip's seed, so Post-Production refines with it.
Post-Production, step by step
Check settings, load the clip, upscale the video latent x2.00 (the trained H3 Latent Upscaler node if installed, else a built-in resize), build the model, sample 3 euler steps (sigmas 0.9035, 0.6316, 0.3158, 0), decode video and audio, create the video. Production mode = the node returns ExecutionBlockers and does nothing.
The UI
Plain ES modules in web/. The slide-out modules are panels built as children of the node's own DOM element and kept in step by timers (background pages run no animation frames). Hidden "ghost" nodes carry the real widgets; the DOM controls mirror them. Widgets are restored by name (properties.ironleo78_values), so a layout change never shifts a saved value. Colours come from one theme table (il78_theme.js), contrast-tested.
Hardcoded on purpose
The diffusion model, both VAEs, the text encoder, attention backend, upscaler, the refine schedule, the Casting LLM and its settings, the LoRa (always off), 24 fps, h264. Fewer knobs, fewer ways to get a bad clip. Model file names are fixed in settings.py, revitalizer.py and ltv.py.
Folder layout
custom_nodes/ironleo78_mtv/
├── __init__.py, mtv.py, revitalizer.py, prompt_generator.py ...
├── presets/ (Casting presets, plain .md)
├── prompt_presets/ (your saved prompts, plain .txt)
├── guide/ (the LLM's system prompts)
├── web/ (the front end)
├── tests/, tools/ (checks and the release script)
└── README.md, ARCHITECTURE.md, CHANGELOG.md, LICENSE
Troubleshooting and dependencies
Troubleshooting
| Problem | What to do |
|---|---|
| Nodes missing | The folder must sit directly in custom_nodes/ with __init__.py inside. Restart ComfyUI, Ctrl+F5, and look in the console for an import error mentioning ironleo78_mtv. |
| Import error about MiniMax H3 | Update ComfyUI: the MTV needs its core MiniMax H3 nodes. |
| A model is "not found" | Put the file in the matching ComfyUI/models/ folder with its exact name (table at the top) and press R. |
| Side panels or buttons look wrong | Use the Nodes 2.0 (Vue) front end and hard-refresh (Ctrl+F5). |
| Panel says "Segments: unknown" | Pick a Custom Audio File (Choose Audio latent = Custom Audio) or a Video File. |
| The node stops and does nothing | You passed the last segment. Use Go, Latest or Reset in the panel. |
| Out of memory | Close other GPU programs, use Test, or retry. The console prints the failed step and a tip. |
| Post-Production list says "none found" | Make a clip, press R, or use Upload. |
| "audio latent is missing" or "conditioning not found" | The three files of a clip must share a name and folder. Upload all three together. |
| Status says upscale: ... built in | The optional upscaler node or its model is not installed; the built-in resize was used. |
| Post-Production does nothing | You are in Production: the node is bypassed. Click Post-Production on the MTV. Press SELECT YOUR LATENT AND CLICK TO REMASTER to run it alone. |
| A phone video lies on its side | The pack reads the phone's rotation and turns it upright. A mirrored video is not un-mirrored. |
| Prompt text looks wrong after changing a preset | Choosing a preset or the mode always shows its example prompt; use Saved prompts to keep yours. |
Dependencies
ComfyUI must be recent enough to include (the pack calls them by name): the MiniMax H3 nodes (comfy_extras.nodes_minimax_h3), LTXVSeparateAVLatent, LTXVConcatAVLatent, CreateVideo, SaveVideo, ModelAttentionBackend (optional), and the core samplers and loaders (RandomNoise, BasicGuider, BasicScheduler, KSamplerSelect, SamplerCustomAdvanced, UNETLoader, CLIPLoader, VAELoader, VAEDecode, VAEDecodeAudio, VAEEncodeAudio, TrimAudioDuration, SolidMask, SetLatentNoiseMask).
Python packages: none to install. torch, torchaudio, numpy, Pillow, PyAV, safetensors and aiohttp ship with ComfyUI. ffmpeg is not needed.
Optional: the Comfyui_Minimax_h3_latent_Upscaler node (MIT, (c) LBH-123-AI) for the trained upscaler in Post-Production:
cd ComfyUI/custom_nodes
git clone https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler.git
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 models: LTX-2.5 is gated on Hugging Face: request access on its page first. Qwen3-VL 4B / 8B encoders do not work as the H3 text encoder (hidden size mismatch).
Run the tests (optional): python tools/run_all.py (uses ComfyUI's own Python and installs jsdom once for the JS tests).
License: MIT (see LICENSE). Changelog: CHANGELOG.md. Architecture notes: ARCHITECTURE.md.
by IronLeo78