DaSiWa MiniMax H3 Workflows v1.4

This is a customized version of the original DaSiWa MiniMax H3 workflow.

Original workflow page

It preserves the original T2VA, I2VA, FL2VA, and REF2VA modes while adding quality-of-life tools for attention selection, clip continuation, seed management, automatic aspect-ratio handling, color matching, previews, and optional latent refinement.

What is new in v1.4

• New unified Attention selector in the Settings panel.

• Improved NativeAddGuide and Masked AV continuation routing.

• Fixed the continuation selector so it appears correctly as a Settings widget.

• Fixed multiple audio/video synchronization and exact-duration issues.

• Masked AV now conforms source audio to the exact video timeline.

• Improved automatic resolution and aspect-ratio handling for continuation clips.

• Improved latent-upscale routing and high-resolution conditioning alignment.

• Removed the standalone external prompt input box.

• Prompt creation is now handled directly through the MiniMax H3 Director prompt builder.

Attention Selector

The Settings panel includes one Attention selector:

ValueBackendDescription0NativeMaximum compatibility and reference quality. Usually the slowest option.1SageAttentionFaster attention with generally similar quality to Native.2Comfy Kitchen AttentionRecommended balance between speed, quality, and compatibility.3Sol-AttnExperimental sparse attention. May provide additional acceleration depending on the GPU.

The selected attention backend is applied to both the FL2VA and REF2VA model paths. Only one backend is active at a time.

Attention requirements

• Native: included with ComfyUI.

• SageAttention: requires a compatible SageAttention and KJNodes installation.

• Comfy Kitchen Attention: requires a recent ComfyUI version.

• Sol-Attn: requires ComfyUI-SolAttn_triton.

Spectrum Acceleration

Spectrum is an optional denoising acceleration for MiniMax H3. It can be used together with the selected Attention backend.

For final-quality comparisons, disable Spectrum if you notice:

• Weaker motion

• Reduced face detail

• Ghosting or temporal artifacts

• Audio/video inconsistencies

Repository: ComfyUI-Spectrum-MiniMax-H3

Clip Continuation

The workflow supports two continuation modes:

NativeAddGuide

Uses a previous video as visual and audio context.

Recommended AddGuide Context Frames:

• 1: fastest, minimal context

• 6: light context

• 12: balanced short context

• 22: recommended balance

• 39: stronger continuity, but slower

Masked AV

Preserves an exact visual and audio prefix from a previous clip while generating the remaining section. It is useful when a more precise AV boundary is required.

Use a source video with synchronized audio and video. NativeAddGuide and Masked AV should not be enabled at the same time.

Repository: ComfyUI-H3-Motion-Context-MultiRef

Automatic Aspect Ratio and Color Match

When continuation is enabled, the workflow automatically derives an output resolution close to the source clip’s aspect ratio while respecting the selected megapixel preset and MiniMax H3 size constraints.

Color Match can reduce visible changes in:

• Temperature

• Saturation

• Brightness

• Contrast

• Overall color balance

Recommended Color Match Strength:

• 0.10–0.15: subtle correction

• 0.20: recommended balance

• 0.25–0.35: stronger correction

Color Match is used only when a continuation mode is active. Disable it if you experience VRAM issues or unwanted color changes.

Optional Latent Upscale Refinement

The optional LBH branch generates the video at a lower resolution, upscales the latent in latent space, and performs a final high-resolution refinement pass.

Example:

16 total steps = 12 base-resolution steps + latent upscale + 4 high-resolution steps

Recommended scales:

• 1.25×: lighter and safer

• 1.50×: recommended starting point

• 2.00×: heavier and more demanding

The workflow uses:

• MinimaxH3LatentUpscaler3D

• BF16 precision

• CUDA processing

• Alignment 32

• Optional temporal chunking

• Automatic model unloading

Repository: Comfyui_Minimax_h3_latent_Upscaler

Model download: Minimax H3 Latent Upscaler

Latent refinement is optional and disabled by default. Results may vary depending on resolution, scale, frame count, GPU, and sampling settings.

Previews and Seed Control

The workflow uses KJNodes ModelPreviewOverride for faster previews.

Repository: ComfyUI-KJNodes

An optional MiniMax H3 TAE preview model can be placed in:

ComfyUI/models/vae_approx/taeh3.safetensors

The workflow also uses rgthree Comfy Seed for easier seed comparison, randomization, locking, copying, and reuse.

Repository: rgthree-comfy

Required Components

• A recent ComfyUI build

• FFmpeg

• ComfyUI-GGUF

• ComfyUI-DaSiWa-Nodes

• ComfyUI-KJNodes

• ComfyUI-H3-Motion-Context-MultiRef

• SageAttention, if using Attention 1

• ComfyUI-SolAttn_triton, if using Attention 3

• ComfyUI-Spectrum-MiniMax-H3, if using Spectrum

• Comfyui_Minimax_h3_latent_Upscaler, if using latent refinement

Recommended Starting Settings

• Attention: 2 — Comfy Kitchen Attention

• Spectrum: enabled for testing; disable for final comparisons if artifacts appear

• AddGuide Context Frames: 22

• Color Match Strength: 0.20 for continuation

• Latent Upscale: disabled initially; test from 1.25× or 1.50×

• NativeAddGuide or Masked AV: use only one at a time

For maximum compatibility, use Native Attention, disable Spectrum, and use conservative resolution and refinement settings.

This workflow is intended for experimentation and benchmarking. Results may vary depending on GPU model, VRAM, CUDA/PyTorch versions, ComfyUI version, resolution, frame count, and enabled optimizations.