velvet-vice-minimax-h3 Introduction
velvet-vice-minimax-h3 is a standalone custom-node package for creating video from an image with MiniMax H3 FL2VA in ComfyUI. It brings model selection, prompting, audio choices, sampling, optional refinement, previews, and finishing controls into one H3-focused workflow.
For you as an AI artist, the aim is to make the creative decisions easier to manage: choose the H3 models you have available, direct the motion and scene, decide whether the result needs sound, and optionally refine or finish the video. The package also includes checks and a live timer to help you follow a render.
The current public workflow and package version is v1.5.0. This is a MiniMax H3-only package, separate from Velvet Vice Zen MiniMax H3.
How velvet-vice-minimax-h3 Works
Think of the workflow as a sequence of creative passes:
- Choose and direct the render. The H3 System Hub and H3 Director bring together the main controls. The I2VA Vision / Prompt Director helps guide the image-to-video result.
- Make the first video pass. MiniMax H3 uses the selected model and settings to generate the initial video. You can choose WITH SOUND or MUTED · VIDEO ONLY.
- Optionally refine the result. Quality Refine runs a second sampling pass before the final video is decoded. You can leave it off for a simpler render or enable it to make a further pass over the generated video.
- Preview and finish. Preview tiers let you balance preview quality against the decoding method available. The output and finishing controls provide additional options before the final result is saved.
The workflow routes between supported native and GGUF H3 model backends and between native and GGUF Qwen text encoders. In practical terms, its selectors are designed to work with the compatible files you already have installed. Selecting an available file does not, by itself, turn on every related option: for example, a detected Turbo LoRA is applied only when Turbo is enabled in the H3 Director.
velvet-vice-minimax-h3 Features
Creative direction and model controls
- H3 System Hub and H3 Director: Central controls for the H3 render, including the integrated Quality Refine settings.
- I2VA Vision / Prompt Director: Controls for directing the image-to-video prompt. Its selected mode persists when you switch workflows.
- Native and GGUF routing: Select from the installed H3 model files listed by the workflow. The Qwen text-encoder selector likewise routes between installed native and GGUF options.
- Persistent H3 and LoRA profiles: Retain your chosen profiles for later use.
Sound and sampling
- WITH SOUND / MUTED · VIDEO ONLY: The sound choice carries through prompting, VAE loading, decoding, and the final output. In muted mode, the audio VAE is skipped.
- Power LoRA AV stack: Provides a LoRA-based path within the H3 workflow.
- Turbo / Distilled LoRA path: Provides separate base and turbo step controls. A compatible Turbo LoRA is strongly recommended by the author because it can substantially reduce sampling cost. Match Turbo Steps to the particular LoRA you downloaded.
Quality Refine / Second Sampler
Quality Refine runs after the first H3 sampling pass and before the final VAE decode. It is OFF by default; when off, the second pass is bypassed rather than run unnecessarily.
| Mode | What it does |
|---|---|
| OFF | Uses the first pass without running a second sampler. |
| LIGHT | Runs a second pass with your chosen Refine Steps and denoise set to 0.12. |
| HIGH | Runs a second pass with your chosen Refine Steps and denoise set to 0.20. |
| CUSTOM | Lets you choose 1–16 Refine Steps and denoise from 0.01–0.35. |
Denoise controls how strongly the refinement pass can change its starting result. For example, LIGHT uses a lower denoise setting than HIGH, while CUSTOM lets you choose a value within the stated range. Refine Steps remain under your direct control in LIGHT, HIGH, and CUSTOM modes.
Two optional controls change how the second pass is prepared:
- Preserve Base Audio: Keeps the audio from the first pass while the second pass refines the video.
- Decode → Upscale → Re-Encode: Prepares a higher-resolution video latent before the second pass, with an adjustable scale. This is an optional refinement path, not a required step for every render.
Previews, finishing, and output
- Live-preview tiers: LOW uses
latent2rgband needs no extra preview decoder. AUTO / MEDIUM uses TAEHV when its optional decoder is installed, otherwise falls back tolatent2rgb. HIGH uses the full MiniMax H3 video VAE, with a fallback chain. - Preflight and Render Timer: The H3 Preflight monitor and live Render Timer help you check the workflow and follow an active render.
- Output and finishing controls: The H3 Output / Finishing Hub and Output Studio include a RIFE / Ghost Analyzer / Temporal Anti-Ghost finishing path and integrated watermark controls.
- Terminal-frame PNG export: The workflow automatically exports a PNG decoded from the actual saved video’s final frame. If that extraction cannot complete, it has a non-fatal tensor fallback.
- Memory cleanup and telemetry: Final cleanup and runtime information are included to help manage and observe the render.
The H3 workflow uses a static Midnight Violet / Obsidian appearance; execution does not recolor its node bars green or teal. Its control surfaces, including H3 POWER LoRA AV, can be resized.
velvet-vice-minimax-h3 Models
This package is focused on MiniMax H3 FL2VA image-to-video. The provided information does not identify a list of separate H3 model releases. Instead, the workflow offers several ways to load or support an H3 render:
- Native or GGUF H3 backend: Choose a compatible installed H3 model from the available-model list. These are backend options within the H3 workflow, not separately named creative models in this documentation.
- Native or GGUF Qwen text encoder: The workflow routes to an installed text-encoder option through one selector.
- Native ComfyUI video and audio VAEs: These handle video and audio decoding as needed by the selected sound mode. The audio VAE is skipped for MUTED · VIDEO ONLY.
- Optional TAEHV preview decoder: The file
models/vae_approx/taeh3_decoder.safetensorsenables the TAEHV preview path used by AUTO / MEDIUM when available. It is not required for LOW previews. - Optional compatible Turbo LoRA: This is an additional LoRA path, not a replacement for enabling Turbo in the H3 Director. Different Turbo LoRAs may call for different Turbo Steps, so use the setting that matches the one you obtained.
Troubleshooting velvet-vice-minimax-h3
A Turbo LoRA says “AVAILABLE/READY,” but Turbo is not being used
- Check whether Turbo is enabled in the H3 Director. “AVAILABLE/READY” means the file was detected, not that it has been applied.
- Check that Turbo Steps matches the Turbo LoRA you downloaded.
- If you prefer a render without Turbo, leave Turbo disabled; H3 also works in Base mode.
Quality Refine does not seem to run
- Check its mode in the H3 System Hub. OFF is the default and bypasses pass 2.
- Select LIGHT, HIGH, or CUSTOM to run refinement.
- Set your desired Refine Steps. In CUSTOM mode, the documented range is
1–16steps and0.01–0.35denoise.
The preview does not use the decoder you expected
- Check the selected preview tier.
- Use LOW if you want the
latent2rgbpath without an additional preview decoder. - For AUTO / MEDIUM, check whether
models/vae_approx/taeh3_decoder.safetensorsis present. Without it, the workflow falls back tolatent2rgb. - HIGH uses the full H3 video VAE when available and also has a fallback chain.
The output has no sound
Check whether MUTED · VIDEO ONLY is selected. That choice is carried through the workflow and skips the audio VAE. To use the sound-enabled path, select WITH SOUND. If you are using Quality Refine and want to keep audio from the first pass, check Preserve Base Audio.
Frequently asked questions
Do I need Quality Refine for every video?
No. It is off by default, and the workflow bypasses the second pass when you leave it off.
What is the difference between LIGHT and HIGH refinement?
Both use the Refine Steps you select. LIGHT uses denoise 0.12; HIGH uses 0.20. CUSTOM gives you control over denoise within the documented range.
Do I need the optional preview decoder?
No. LOW needs no extra decoder, and AUTO / MEDIUM falls back to latent2rgb when TAEHV is unavailable.
Is this the same package as Velvet Vice Zen MiniMax H3?
No. This standalone H3 package is kept separate from the Velvet Vice Zen MiniMax H3 package.
Learn More about velvet-vice-minimax-h3
- Compatible Turbo LoRAs: MiniMax H3 Turbo LoRAs on Civitai (https://civitai.red/models/2837571/minimax-h3-turbo-loras?modelVersionId=3275758). Check which LoRA you choose so you can match its Turbo Steps in the H3 Director.
- Workflow version: Use the current v1.5.0 public workflow with the v1.5.0 package when following this documentation.
