MiniMax-H3 INT8 FLF, R2V & TTS Workflows Collection

Optimized ComfyUI workflows for MiniMax-H3, featuring INT8 inference, Sage-Attention, Spectrum, Lightx2v LoRA, Turbo-LoRA, Motion Context, Latent Upscaling, and high-speed Text-to-Speech.

Workflows

Base model:

Comfy-Org/MiniMax-H3
https://huggingface.co/Comfy-Org/MiniMax-H3

---

Features

INT8

INT8 MiniMax-H3 inference for:

• Reduced VRAM usage
• Faster inference
• Improved memory efficiency

Sage-Attention

Optimized attention implementation for faster MiniMax-H3 inference and improved generation throughput.

Spectrum

ComfyUI-Spectrum-MiniMax-H3
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

Optimized sampling and MiniMax-H3 inference acceleration.

Motion Context

ComfyUI-H3-Motion-Context
https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

Provides additional temporal motion conditioning for FLF and R2V workflows.

Lightx2v LoRA

lightx2v/Minimax-h3-Turbo
https://huggingface.co/lightx2v/Minimax-h3-Turbo

Acceleration LoRA for reduced sampling steps and faster MiniMax-H3 video generation.

Turbo-LoRA

MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Experimental low-step acceleration. Output quality may vary depending on LoRA strength and sampling settings.

R2V Reference LoRA

The R2V workflow requires the following reference LoRA:

minimax_h3_ref_lora_rank_256_bf16.safetensors

https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

This LoRA provides the reference-image conditioning required by the R2V workflow.

Latent Upscaler (Experimental)

ComfyUI-MiniMaxH3_LatentUpscaler
https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

Post-generation latent video upscaling for higher resolution and improved detail.

Note: The latent upscaling function is currently not working reliably. If using the workflow, set Enable 2Pass to False.

Face Refine

ComfyUI-H3-FaceRefine
https://github.com/Carasibana/ComfyUI-H3-FaceRefine

Provides automatic face refinement for generated video, similar to a Face Detailer workflow.

Features:

• Automatic face detection
• Face refinement
• Facial detail enhancement
• Improved facial consistency
• Video face correction

Text-to-Speech

MiniMax-H3 can also be used as a dedicated Text-to-Speech engine.

The TTS workflow generates audio only, bypassing video generation entirely. This significantly reduces processing requirements and enables extremely fast speech synthesis.

---

Workflow Details

First-Last-Frame (FLF)

MiniMax_int8_I2V-javanoYYMM.x.json

Generates controlled transitions between a first and last frame.

Features:

• First + Last Frame conditioning
• Motion Context
• INT8
• Sage-Attention
• Spectrum
• Lightx2v LoRA
• Optional Turbo-LoRA
• Latent Upscaler

Recommended for cinematic transitions, character animation, camera movement, transformations, and keyframe-based video generation.

---

Reference-to-Video (R2V)

MiniMax_int8_R2V-javanoYYMM.x.json

Generates video from a reference image with dedicated reference LoRA and Motion Context conditioning.

Features:

• Reference image conditioning
• Reference LoRA
• Character/appearance consistency
• Motion Context
• INT8
• Sage-Attention
• Spectrum
• Lightx2v LoRA
• Optional Turbo-LoRA
• Latent Upscaler

Recommended for:

• Character videos
• Consistent subjects
• Appearance preservation
• Image-driven animation
• Reference-guided video generation

Required R2V LoRA

minimax_h3_ref_lora_rank_256_bf16.safetensors

Download from:

https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

Place the LoRA in the appropriate ComfyUI LoRA directory and load it in the R2V workflow.

---

Face Refine

MiniMax_int8_FR-javano2608.1.json

A dedicated face refinement workflow for MiniMax-H3 generated video.

Uses:

ComfyUI-H3-FaceRefine
https://github.com/Carasibana/ComfyUI-H3-FaceRefine

The workflow automatically detects and refines faces in the generated video, providing a Face Detailer-like correction process.

Recommended for:

• Facial detail enhancement
• Correcting facial artifacts
• Improving face consistency
• Refining generated characters

---

Text-to-Speech (TTS)

MiniMax_int8_TTS-javanoYYMM.x.json

A dedicated Text-to-Speech workflow using MiniMax-H3.

The workflow outputs audio only, without generating video frames. By bypassing the video generation pipeline, it enables extremely fast speech synthesis with significantly lower processing requirements.

Features:

• Text-to-Speech using MiniMax-H3
• Audio-only output
• INT8 inference
• Fast speech generation
• ComfyUI workflow integration

Recommended for:

• Voice generation
• Dialogue creation
• Voice-over production
• Rapid audio prototyping
• Generating speech for video production

---

Required Custom Nodes

• ComfyUI-Spectrum-MiniMax-H3
  https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

• ComfyUI-H3-Motion-Context
  https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

• ComfyUI-MiniMaxH3_LatentUpscaler
  https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

• ComfyUI-H3-FaceRefine
  https://github.com/Carasibana/ComfyUI-H3-FaceRefine

Install through ComfyUI Manager → Install Missing Custom Nodes.

---

Required Models

Main Model

Comfy-Org/MiniMax-H3
https://huggingface.co/Comfy-Org/MiniMax-H3

R2V Reference LoRA

minimax_h3_ref_lora_rank_256_bf16.safetensors
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

Required for the R2V workflow.

Lightx2v LoRA

lightx2v/Minimax-h3-Turbo
https://huggingface.co/lightx2v/Minimax-h3-Turbo

Used for accelerated video generation with reduced sampling steps.

Optional Turbo-LoRA

MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Used for experimental low-step acceleration.

---

Installation

1. Update ComfyUI and ComfyUI-Manager.

2. Install the required custom nodes listed above.

3. Download Comfy-Org/MiniMax-H3.

4. For R2V, download:

   • minimax_h3_ref_lora_rank_256_bf16.safetensors

5. Download lightx2v/Minimax-h3-Turbo if using Lightx2v acceleration.

6. Download MiniMax-H3-Turbo-Lora-ComfyUI if using Turbo-LoRA.

7. Load the desired workflow:

   • MiniMax_int8_I2V-javanoYYMM.x.json — FLF
   • MiniMax_int8_R2V-javanoYYMM.x.json — R2V
   • MiniMax_int8_FR-javano2608.1.json — Face Refine
   • MiniMax_int8_TTS-javanoYYMM.x.json — TTS

---

Recommended Configuration

High-Speed Configuration:

INT8 + Sage-Attention + Lightx2v LoRA or Turbo-LoRA

The optimal configuration may vary depending on the workflow, GPU, sampling settings, LoRA strength, and desired output quality.

---

Notes

• FLF = First-Last-Frame.
• R2V = Reference-to-Video.
• FR = Face Refine.
• TTS = Text-to-Speech.
• Motion Context supports both FLF and R2V.
• The R2V workflow requires minimax_h3_ref_lora_rank_256_bf16.safetensors.
• The TTS workflow outputs audio only and does not generate video.
• Turbo-LoRA is experimental.
• Sage-Attention and Spectrum require compatible environments.
• Combining multiple acceleration methods may require tuning of sampling steps and LoRA weights.
• The Latent Upscaler is currently not fully reliable. Set Enable 2Pass to False when using it.
• If audio is not generated correctly, try lowering the shift_audio value of the ModelSamplingMiniMaxH3 node inside the INT8 Model Loader subgraph from 4.00 to 3.00.

---

Additional Resources

MiniMax-H3
https://huggingface.co/Comfy-Org/MiniMax-H3

R2V Reference LoRA
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

Spectrum MiniMax-H3
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

H3 Motion Context
https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

H3 Face Refine
https://github.com/Carasibana/ComfyUI-H3-FaceRefine

Lightx2v LoRA
https://huggingface.co/lightx2v/Minimax-h3-Turbo

MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

MiniMax-H3 Latent Upscaler
https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

---

License

Please follow the respective licenses of MiniMax-H3, ComfyUI, the listed custom nodes, LoRAs, and other third-party components used by these workflows.