MiniMax Music 3 Full MV Pipeline — For Patreon Supporters (Patreon)
Downloads
Content
Hey everyone.
If you watched the MiniMax Music 3 video https://www.patreon.com/aifuturetech/posts/minimax-music-3-166693117 , you saw the basic text-to-music workflow and the H3 audio lipsync demo. This post has the production version — the full pipeline from a single idea to a finished music video, with the M3 SongPlanner node that writes the caption and lyrics for you.
The MV demo, not prefect, wait for you guys to do a better one :)
MiniMax-Music3
https://modelscope.cn/models/MiniMax/MiniMax-H3
https://huggingface.co/MiniMaxAI/MiniMax-Music3
https://minimax-ai.github.io/music3-demo/
MiniMax-Music3 for ComfyUI
https://huggingface.co/Comfy-Org/MiniMax-Music-3/
ComfyUI MiniMax M3 SongPlanner
https://github.com/benjiyaya/ComfyUI-MiniMax-M3-SongPlanner
Minimax-H3-Prompt-AgentSkill
https://github.com/benjiyaya/Minimax-H3-Prompt-AgentSkill
ComfyUI-H3-VisionPromptor
https://github.com/benjiyaya/ComfyUI-H3-VisionPromptor
----------------------------------------
WHAT'S INCLUDED
----------------------------------------
1. M3 SongPlanner — Music 3 With Music Planner
.....................................
File: Minimax Music 3 - With Music Planner.json
This is the workflow from the video where you type one idea and get a complete song. It adds the M3SongPlanner custom node in front of the standard Music 3 text encoder, so you never write a caption or lyrics by hand again.
The pipeline:
- M3SongPlanner node takes your idea, genre, vocal config, language, and target duration
- It uses a Qwen3 VL 4B text encoder locally to generate the structured caption and section-tagged lyrics
- Output connects directly to the MiniMaxMusic3TextEncode node
- Music 3 DiT (FP16/FP32), text encoder, and DAV VAE loaders
- KSampler (euler, 30 steps, CFG 1.7), tiled audio decode for VRAM safety
- ShowText nodes so you can review the generated caption and lyrics before you queue the audio
The example loaded in the workflow: a dark K-pop song about love and loss, 60 seconds, female vocals, English, seed 3877326292, temperature 0.8, top-p 0.95.
The key difference from the free version: the planner node writes everything for you. You give it an idea and it produces the full three-section caption (global metadata, vocal details, arrangement) plus lyrics with [Intro] [Verse] [Chorus] [Bridge] [Outro] tags.
Key settings:
- temperature: 0.8 — good balance of creativity and coherence
- top_p: 0.95 — keeps the lyric generation stable
- max_duration: set this 50% longer than your target (60s lyrics → 90s max duration)
Dependencies inside the workflow:
- models/diffusion_models/minimax_music3_dit_fp16.safetensors (or fp32)
- models/text_encoders/minimax_music3_text_encoder_pruned_int8_convrot.safetensors
- models/vae/minimax_music3_dav.safetensors
- models/text_encoders/qwen3vl_4b_bf16.safetensors (for the planner)
- ComfyUI-MiniMax-M3-SongPlanner custom node pack
2. Krea 2 Character Sheet Generator
.....................................
File: Krea2_CharacterSheet-20260814.json
Creates the character reference sheet used for the MV — 44 nodes that generate a multi-pose character sheet in one pass.
The pipeline:
- Krea 2 Turbo checkpoint with the QuadView LoRA (krea2_QuadView_krea2_v1.safetensors at 1.0) and realism LoRA (krea2_Krea2-realism-V2.safetensors at 1.0)
- Three KSamplers: 10-step euler for the initial sheet, 8-step euler at CFG 0.3 for the refine pass, 8-step res_multistep for the final output
- Character prompt input that describes the character (appearance, role, wardrobe)
- Refine prompt toggle to regenerate weak panels
This is the workflow I used to create the K-pop style character sheet you see in the video. Feed it a character description and it outputs a consistent multi-angle reference that H3 can use for identity across all six scenes.
Key settings:
- QuadView LoRA at 1.0 — this is what forces the 4-panel character sheet layout
- Realism LoRA at 1.0 — pushes the output toward photorealistic
- First KSampler: euler, 10 steps, CFG 1 (initial layout)
- Second KSampler: euler, 8 steps, CFG 0.3 (refine pass)
- Output at 1MP is the sweet spot for character sheets
Dependencies inside the workflow:
- models/checkpoints/krea2_turbo_fp8_scaled.safetensors (or krea2_raw_bf16)
- models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors (type krea2)
- models/vae/qwen_image_vae.safetensors
- models/loras/krea2/QuadView_krea2_v1.safetensors
- models/loras/krea2/Krea2-realism-V2.safetensors
3. H3 Music Video Pipeline — LTX 2.5 Upsampling
.....................................
File: MiniMax-H3 R2V - MusicVideo Pipeline (Advance) -20260814.json
The main MV pipeline — 64 nodes. Generates each scene with MiniMax H3 reference-to-video (image + audio), then upscales and frame-interpolates to a finished 60fps music video clip.
The pipeline:
- MiniMaxH3ReferenceToVideo with the ref2va int8 model and the turbo 4-step LoRA at 0.7
- Qwen3VL 32B text encoder, H3 video VAE + audio VAE
- Reference image (your character sheet) and reference audio (your Music 3 song)
- SeedVR2VideoUpscaler (720 base) + RTXVideoSuperResolution (1.5x, ULTRA preset)
- FILM frame interpolation to 2x (60fps from 30fps)
- VHS_VideoCombine outputs for each scene
- Audio passthrough preserved through the whole chain
The upscaling path in this version uses the LTX 2.5 family: LTX 2.5 latent spatial upscaler and LTX 2.5 audio VAE for the upscale stage.
Key settings:
- H3 turbo LoRA at 0.7 — the sweet spot for 4-step sampling without losing motion quality
- SeedVR2 at 720 base resolution, then RTX 1.5x ULTRA on top
- FrameInterpolate 2x with FILM
- 16:9 widescreen, 0.6MP base resolution
- euler sampler
Dependencies inside the workflow:
- models/unet/minimax-h3/minimax_h3_ref2va_int8_convrot.safetensors
- models/loras/minimax-h3/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
- models/text_encoders/minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors
- models/vae/minimax-h3/minimax_h3_video_vae_fp16.safetensors + minimax_h3_audio_vae_fp32.safetensors
- models/upscale_models/ltx-2.5/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
- models/vae/ltx-2.5/ltx-2.5-audio-vae-bf16.safetensors
- models/text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
- FILM frame interpolation model
4. H3 Music Video Pipeline — LTX 2.3 Upsampling (ALTERNATIVE)
.....................................
File: MiniMax-H3 R2V - 2.3 MusicVideo Pipeline (Advance) -20260814.json
The same MV pipeline, but the upsampling stage uses the LTX 2.3 family instead of LTX 2.5 — 73 nodes. Identical H3 generation stage, different upscaler chain.
The pipeline:
- Same MiniMaxH3ReferenceToVideo + ref2va + turbo 4-step LoRA at 0.7
- Same SeedVR2 + RTX upscale chain
- Same FILM 2x frame interpolation
- The upscale stage instead loads LTX 2.3: ltx-2.3-22b-dev checkpoint, LTX 2.3 video VAE, Gemma 3 12B text encoder, and the distilled LoRA at 1.0
Key settings:
- LTX 2.3 distilled LoRA at 1.0 for the upscale pass
- Base resolution 0.4MP (lower than the 2.5 version) because the 2.3 upscaler handles the lift differently
- Everything else matches the 2.5 version: euler, turbo LoRA 0.7, SeedVR2 720, RTX 1.5x ULTRA, FILM 2x
WHICH ONE SHOULD YOU USE?
I built both versions because they genuinely produce different outcomes, and there's no single winner:
- LTX 2.5 upsampling: cleaner details and better structure on close-up scenes, slightly different motion feel. Best if your MV has lots of face close-ups (like the "Still Here" outro scene in the demo).
- LTX 2.3 upsampling: faster overall run, smoother on fast camera moves, different texture on the upscale. Best if your MV is mostly wide and medium shots.
Try the same scene through both and pick the one that fits your footage. That's exactly what I do when I cut a new MV. The settings differ slightly between them (0.6MP vs 0.4MP base), so keep that in mind when comparing.
----------------------------------------
QUICK START
----------------------------------------
1. Import the four workflow files into ComfyUI (drag and drop).
2. Install the custom node packs listed above via ComfyUI Manager.
3. Check every loader node points to a model you actually have.
4. Start with the M3 SongPlanner workflow: type your idea in the M3SongPlanner node, review the generated caption and lyrics in the ShowText nodes, then queue.
5. Set max_duration to about 1.5x your target song length (60s lyrics → 90s max).
6. Run the Krea 2 character sheet workflow with your character description to get a reference sheet.
7. In the H3 MV pipeline, load the character sheet as the reference image and your generated song as the reference audio.
8. Copy the scene prompt from the video's plan into the text prompt node, set duration, queue.
9. After all scenes render, compare the LTX 2.3 and LTX 2.5 upscale outputs on the same scene and pick your favorite.
10. Stitch the final scenes in your editor.
----------------------------------------
TIPS
----------------------------------------
1. Max duration math: 60 seconds of lyrics needs 90 or more seconds of max duration. The intro music, gaps between sections, and outro all consume time. I learned this the hard way when the ending got cut off on the official site.
2. Share the seed between the Music 3 text encoder and the KSampler. Connected seeds give you consistent inversion across runs, which matters when you regenerate a scene to fix a detail.
3. The M3SongPlanner temperature at 0.8 is the sweet spot. Crank it higher and lyrics get chaotic; lower and they get generic. Top-p 0.95 keeps the structure tags clean.
4. H3 turbo LoRA at 0.7, not 1.0. Full strength makes motion aggressive and can break character consistency across scenes. 0.7 keeps the 4-step speed without the motion artifacts.
5. Write the scene prompts for the audio, not the image. The H3 reference-to-video prompt should describe what happens in the music (lip-sync lines, emotional beats) because that's what the model is synchronizing to.
6. Review every generated scene before stitching. Audio sync slips happen on some scenes — regenerate those individually rather than fixing in post. Much faster.
7. Character consistency is automatic with the same reference sheet. As long as every scene uses the same character image and the turbo LoRA stays at 0.7, the identity holds across all six scenes.
8. The 2.3 vs 2.5 upscaler choice is per-scene, not per-project. Close-ups lean 2.5, wide shots lean 2.3. Mix them in one MV if you want.
----------------------------------------
COMMON ISSUES
----------------------------------------
Song cuts off before the lyrics finish
→ The max duration is too close to the lyrics length. Set it 50% longer. 60s of lyrics needs 80-90s of max duration.
Caption and lyrics look generic
→ Lower the M3SongPlanner temperature to 0.7-0.8 and check top-p at 0.95. Also give the idea field more specifics: mood, setting, instruments, not just genre.
Character looks different between scenes
→ The turbo LoRA strength drifted or the reference sheet has inconsistent poses. Keep the LoRA at 0.7 and use the QuadView character sheet (multi-pose) as reference, not a single image.
Scene audio is out of sync
→ Happens on some scenes with the ref2va model. Regenerate that scene with the same seed and prompt. If it persists, shorten the scene duration to 10 seconds.
VRAM overflow on the upscale stage
→ Drop the base resolution from 0.6MP to 0.4MP in the Resolution Selector, or run SeedVR2 at 540 instead of 720. The RTX upscaler needs headroom.
LTX 2.5 upscaler loads but errors
→ Check you have the LTX 2.5 audio VAE and Gemma 4 text encoder paths correct. The 2.5 chain needs all three files (transformer, audio VAE, text encoder), not just the upscaler.
----------------------------------------
The workflow files are attached. Install the custom nodes first, then check the model paths. Everything else is in the nodes.
Questions? Drop them in the comments and I will help you sort it out.
Thanks for supporting the channel. You guys make these deep dives possible.
— Benji