VELVET VICE — MiniMax H3
Velvet Vice MiniMax H3 image-to-video custom nodes with Native/GGUF routing, Quality Refine second sampling, persistent prompting, live preview and finishing tools.
Nodes (37)
Skip the audio VAE entirely when the final cut is MUTED
The last word on whether your H3 video actually gets sound
How H3 keeps pass-1 audio while pass 2 polishes the video
The node that keeps H3's pass-1 audio when your refine pass rewrites the video
One text-encoder loader that swallows native and GGUF H3 files alike
The lazy bouncer that stops Quality Refine's conditioning from ever running when it's off
The one panel where an H3 render actually gets designed
See the exact prompt H3 is about to eat, before you burn an hour
Finding the ghost frames that frame interpolation leaves behind
A RAM/VRAM breadcrumb between the nodes that eat your card
An exact Lanczos resize for H3 refine frames — no guessing, no cropping
The lazy switch that makes Quality Refine free when it's off
Watch an H3 render cook in real time without trusting the pixels
The same live preview, positioned where the graph actually needs it
A node whose only job is to say the preview pipeline is armed
The switchman that unloads native H3 before a GGUF render takes the track
Kick the vision LLM out of VRAM before the 33B model moves in
The LoRA loader that's honest about being optional
One finishing panel for preview, RIFE, ghost repair and watermarks
Where H3 frames and native audio become the actual video file
Stacking H3 LoRAs in order, with FULL / VIDEO / AUDIO routing
A pre-flight checklist you can run before a 33B render commits
Save your whole H3 setup as a named profile that survives restarts
Turn a source image and a one-line idea into a full H3 prompt
Re-packaging H3's [video, audio] latent so the second pass is even legal
The 32-pixel grid matcher behind H3's re-encode refine path
A stopwatch for your render, not another model loader
Let the node read your reference image's shape, then hand H3 a resolution it can actually digest
The last node in the graph, there to make sure the next render doesn't start in a memory hole
The bouncer that stops a 40GB model from loading into a machine with no RAM left
The single panel that runs your whole MiniMax H3 backend — and the second sampler hiding inside it
Cleaning up the double-exposure smears that frame interpolation leaves behind
One switch flips H3 between 20-step base quality and 8-step turbo speed — safely
The H3 loader that finds your model no matter which folder you buried it in
One loader, two VAEs — because H3 doesn't decode pictures and sound with the same autoencoder
The internal glue that stops your VAE loaders from crossing a subgraph boundary and exploding
The watermark worker you'll probably never click — and that's the point
VELVET VICE — MiniMax H3
Standalone Velvet Vice custom-node package for MiniMax H3 FL2VA image-to-video in ComfyUI.
Scope
This repository is MiniMax H3-only. The public node IDs, workflow metadata, UI, guides, HTTP routes and Comfy Registry package use the standalone ID velvet-vice-minimax-h3.
The current public workflow and package version is v1.5.0.
Main features
- Native / GGUF H3 backend routing with complete installed-model lists
- Automatic native/GGUF Qwen text-encoder routing in one selector
- Native ComfyUI video/audio VAE routing; the audio VAE is skipped in
MUTED · VIDEO ONLY - H3 System Hub and H3 Director
- Quality Refine / Second Sampler integrated into H3 System Hub
- Quality Refine is OFF by default and uses a true lazy bypass when disabled
- Refine Steps remain directly user-controlled in LIGHT, HIGH and CUSTOM modes
- LIGHT uses denoise
0.12; HIGH uses0.20; CUSTOM exposes denoise0.01–0.35 - Optional Preserve Base Audio keeps pass-1 audio while pass 2 refines the video latent
- Optional Decode → Upscale → Re-Encode path before Pass 2 with adjustable scale
- H3-native I2VA Vision / Prompt Director with persistent mode across workflow switches
- VELVET VICE H3 Preflight and live Render Timer monitors
WITH SOUND/MUTED · VIDEO ONLYpropagated through prompting, VAE loading, decoding and final output- Persistent H3 and LoRA profiles
- Power LoRA AV stack
- Turbo / Distilled LoRA path with separate base/turbo steps
- LOW / MEDIUM / HIGH live-preview tiers
- H3 Output / Finishing Hub and Output Studio
- RIFE / Ghost Analyzer / Temporal Anti-Ghost finishing path
- Watermark controls integrated into the H3 Finishing Hub
- Final memory cleanup and runtime telemetry
- Automatic PNG export decoded from the actual saved video's terminal frame, with a non-fatal tensor fallback
- Static Midnight Violet / Obsidian styling across the H3 workflow; execution no longer recolors node bars green/teal
- Freely resizable control surfaces, including the H3 POWER LoRA AV panel
- Strict runtime isolation from the separate Velvet Vice Zen MiniMax H3 package
Quality Refine / Second Sampler
Quality Refine runs after the primary H3 sampling pass and before final VAE decode.
- OFF — default; pass 2 is not executed
- LIGHT — user-selected Refine Steps, denoise
0.12 - HIGH — user-selected Refine Steps, denoise
0.20 - CUSTOM — user-selected Refine Steps
1–16, denoise0.01–0.35 - Preserve Base Audio ON — final audio remains from pass 1 while pass 2 refines video
The second pass uses the same effective H3 model/conditioning with res_multistep + simple. An optional Decode → Upscale → Re-Encode path can prepare a higher-resolution video latent before Pass 2 while rebuilding the H3 [video, audio] latent structure.
Strongly recommended — Turbo LoRA
MiniMax H3 works in Base mode, but a compatible Turbo LoRA is strongly recommended because it can reduce sampling cost dramatically.
Turbo LoRAs on Civitai:
https://civitai.red/models/2837571/minimax-h3-turbo-loras?modelVersionId=3275758
Turbo LoRA AVAILABLE/READY means that the file was detected. The LoRA is only applied when Turbo is enabled in the H3 Director. Match Turbo Steps to the LoRA you downloaded.
Optional live-preview decoder
Optional file:
models/vae_approx/taeh3_decoder.safetensors
Preview tiers:
- LOW — latent2rgb; no extra decoder required
- AUTO / MEDIUM — TAEHV when installed, otherwise latent2rgb fallback
- HIGH — full MiniMax H3 video VAE, with fallback chain
Installation
Comfy Registry / Manager
Install:
velvet-vice-minimax-h3
Then restart ComfyUI and load the current v1.5.0 Civitai workflow.
Manual GitHub installation
Clone or copy this repository into:
ComfyUI/custom_nodes/velvet-vice-minimax-h3
Restart ComfyUI afterwards.
Registry
- Publisher:
velvet-vice - Node ID:
velvet-vice-minimax-h3 - Version:
1.5.0 - Display name:
VELVET VICE — MiniMax H3