Share Your Creations
I would be really happy to see what you create with VELVET VICE — MiniMax H3.
If you make videos with this workflow, feel free to share them on Civitai and show how you use the system in your own way.
Your creations, experiments and results are always welcome.
I’m looking forward to seeing what you make with it.
VELVET VICE — MINIMAX H3
MiniMax H3 I2V Workflow
VELVET VICE — MiniMax H3 is a complete ComfyUI workflow for MiniMax H3 image-to-video generation, built around a guided high-end production pipeline.
The workflow combines model management, H3-native vision prompting, Native/GGUF support, audio/video generation, Turbo LoRA acceleration, live preview, profiles and finishing tools inside one coordinated system.
Main Features
Native and GGUF MiniMax H3 support
Separate Native / GGUF model selection with persistent settings
H3 Video VAE + Audio VAE routing
WITH SOUND and MUTED · VIDEO ONLY generation
Integrated H3 Vision / Prompt Director
Manual, Standard Vision, Adult Assisted and Adult Full Auto modes
Native H3 I2VA prompt structure with first-frame anchoring
Integrated H3 Profile Manager
Power LoRA AV with independent Video, Audio and CLIP routing
Dedicated Turbo / Distilled LoRA control
Separate Base and Turbo sampling steps
Adaptive Live Preview with LOW / AUTO / MEDIUM / HIGH quality modes
Automatic verified TAEHV preview decoder support
Integrated Preflight System Check
Live Render Timer
Optional Temporal Anti-Ghost
Optional RIFE 48 FPS interpolation
Integrated watermark and output controls
Final H3 Output Studio with render-stage monitoring
NVENC / CPU H.264 output handling
Automatic final
_LAST_FRAME.pngexportFinal memory cleanup
Multilingual PDF documentation included
The complete workflow remains organized as a guided:
SYSTEM → SOURCE → DIRECTOR → PROMPT → MODEL / TURBO → RENDER CORE → FINISHING
What's New in v1.5.0
Quality Refine / Second Sampler added directly to the H3 System Hub.
Quality Refine is OFF by default and uses a true lazy bypass, so the second sampler does not run unless enabled.
LIGHT / HIGH / CUSTOM refine modes with freely adjustable Refine Steps.
Optional Decode → Upscale → Re-Encode refinement before the second sampling pass.
Proper reconstruction of the MiniMax H3 video + audio latent structure before Pass 2.
Preserve Base Audio option keeps the original first-pass audio while refining the video.
H3 Vision / Prompt Director mode now persists when switching between workflows.
Improved manual node resizing, including resizable H3 control panels.
Fixed layout overlaps around the H3 Engine / Live Preview area.
Updated Midnight Violet / Obsidian UI with static node styling and no unwanted green/teal execution recoloring.
Quality Refine can noticeably improve sharpness and fine detail, especially at 720p, while the original single-pass workflow remains available simply by leaving Quality Refine disabled.
ComfyUI Manager Installation — Recommended
VELVET VICE — MiniMax H3 is officially available through the Comfy Registry.
Install directly through ComfyUI Manager:
velvet-vice-minimax-h3
A Registry installation uses the official release version and allows ComfyUI Manager to handle installation, updates and uninstall normally.
Only one copy of the package should exist inside custom_nodes.
Description
## v1.5.0 — Quality Refine, Re-Encode Refinement, Persistence & UI Polish
### New — Quality Refine / Second Sampler
- Added an optional Quality Refine second sampling pass for improved sharpness, fine detail and overall image clarity
- Quality Refine is integrated directly into the H3 System Hub
- Added a dedicated ON / OFF switch
- When disabled, the second sampler is completely bypassed and only the normal H3 sampling pass runs
- Refine Steps are freely adjustable by the user
- Added LIGHT, HIGH and CUSTOM refine modes
- LIGHT uses a conservative 0.12 denoise preset
- HIGH uses a stronger 0.20 denoise preset
- CUSTOM allows manual denoise control from 0.01 – 0.35
- Added optional Preserve Base Audio to keep the original first-pass audio while refining the video latent
- The second pass uses res_multistep + simple
### New — Optional Decode → Upscale → Re-Encode Refinement
- Added an optional Decode → Upscale → Re-Encode path before the second sampling pass
- Allows the first-pass result to be decoded, resized and encoded back into the H3 latent before refinement
- Upscale factor can be adjusted by the user
- The H3 [video, audio] latent structure is rebuilt correctly before Pass 2
- Audio can remain untouched through Preserve Base Audio
### Improved Image Quality
- Noticeably improved detail retention and clarity, especially at resolutions such as 720p
- Reduced the soft / washed-out appearance that could remain after the primary H3 sampling pass
- Fine structures, textures and edges remain more defined
- Refine remains optional so users can choose between maximum speed and additional quality
### H3 Vision / Prompt Director
- Fixed the Prompt Director mode resetting to MANUAL after switching between workflows
- Selected modes such as STANDARD VISION, ADULT ASSISTED and ADULT FULL AUTO now persist correctly
### UI / Design
- Restored a consistent Midnight Violet / Obsidian design across the H3 workflow
- Removed unwanted green / teal execution coloring from node headers and status bars
- Prevented dynamic render-state colors from breaking the unified Velvet Vice visual design
- Quality Refine controls are now part of the System Hub instead of appearing as a separate floating control node
- Fixed manual resizing for DOM-based H3 control panels
- H3 POWER LoRA AV can now be resized correctly using the visible bottom-right resize handle
### Release Synchronization
- Civitai workflow/package version: v1.5.0
- GitHub package version: v1.5.0
- Comfy Registry package version: v1.5.0
- ZIP version, workflow version, GitHub version and Registry version are now fully synchronized
- Canonical Registry / Manager package ID remains:
velvet-vice-minimax-h3
- Preserved Native / GGUF routing, H3 audio/video generation, Turbo / Distilled LoRA, Power LoRA AV, profiles, live preview, Temporal Anti-Ghost, RIFE interpolation, watermarking and finishing tools