MiniMax H3 Huggingface: Setup Guide & GGUF Workflows - Opensource

MiniMax H3 Huggingface: Setup Guide & GGUF Workflows

Learn how to download MiniMax H3 models from Hugging Face, set up ComfyUI workflows, and optimize GGUF generation for text-to-video and image-to-video.

2026-08-10
MiniMax H3 Wiki Team
Quick Guide
  • MiniMax H3 Huggingface: The primary hub for downloading GGUF, INT8, and FP8 model versions for local generation
  • Turbo LoRA: Dramatically reduces inference time, bringing 5-second video generation down to under 5 minutes
  • GGUF vs INT8: GGUF models are highly recommended for consumer GPUs (like RTX 3080) to avoid stalling and memory errors
  • ComfyUI Workflows: Supports text-to-video, image-to-video, first-last frame, and reference-to-video pipelines
  • Sage Attention: Essential optimization tool that must be enabled alongside Torch Compile for maximum speed

MiniMax H3 Huggingface Models Overview

Hugging Face has become the central repository for the MiniMax H3 ecosystem, hosting a wide variety of quantized models, text encoders, VAEs, and community-created LoRAs. Navigating these files and knowing exactly which ones to download for your specific hardware is the first step toward successful local video generation.

Video Highlights:

  • How to switch from INT8 to GGUF models to prevent generation stalling
  • Step-by-step breakdown of Turbo LoRA installation and configuration
  • Full workflow demonstrations for text-to-video, image-to-video, and reference-to-video
  • Best practices for Sage Attention and Torch Compile on RTX 3080

When browsing MiniMax H3 files on Hugging Face, you will encounter several model types. The two primary models required for most workflows are the First Frame Last Frame (FFLF) model and the Reference to Video (Ref2V) model.

Model Selection Matters

Do not use the Reference to Video model for standard image-to-video or text-to-video tasks. The Ref2V model is strictly designed for handling multiple reference inputs (images, audio, and video). For standard generations, always load the First Frame Last Frame model.

Available GGUF Quantization Levels

Choosing the right quantization (Q-level) depends entirely on your GPU's VRAM. Pushing for a higher quality model on insufficient VRAM will cause ComfyUI to stall before reaching the sampler or throw an Out of Memory (OOM) error.

Model TypeRecommended QuantizationTarget VRAMNotes
FFLF ModelQ4_K_M10GB - 12GBBest balance of speed and quality for RTX 3080
FFLF ModelQ5 / Q816GB+Higher quality, significantly longer generation times
Ref2V ModelQ4_K_M10GB - 12GBRequired for multi-reference identity consistency
Text EncoderQ4 GGUFAllSpeeds up text encoding without noticeable quality loss
Audio/Video VAEOriginal FPAllDo not quantize the VAE; use original files

ComfyUI Installation & Environment Setup

Before downloading models from Hugging Face, you must ensure your ComfyUI environment is fully optimized for MiniMax H3. Standard installations often lack the necessary attention mechanisms required to run these heavy video models efficiently.

Essential Dependencies

To run MiniMax H3 smoothly, you must install Sage Attention and Triton. These libraries handle optimized memory allocation and drastically reduce the time it takes to process each denoising step.

Core Environment Requirements

ComponentRequirementPurpose
ComfyUILatest VersionCore interface for node-based generation
Sage AttentionEnabledAccelerates attention mechanisms, critical for video
TritonInstalledWorks alongside Sage Attention for Windows/Linux
Torch CompileSet to TrueOptimizes PyTorch graph execution; prevents OOM errors
EFF Sage AttentionConditionalEnable if Sigma Shift exceeds 0.4; disable otherwise

Turbo LoRA & Speed Optimization Workflow

The introduction of the Turbo LoRA has fundamentally changed how MiniMax H3 performs on consumer hardware. Previously, generating a 5-second video could take over an hour. With the Turbo LoRA and proper GGUF quantization, that same generation can be completed in roughly 4 minutes and 30 seconds.

Torch Compile is Mandatory

If you encounter an "Out of Memory" error during image-to-video or text-to-video generation, check your Torch Compile setting. Forgetting to set Torch Compile to True is the most common cause of OOM crashes during the sampling phase.

Understanding the Turbo LoRA Setup

While the Turbo LoRA documentation might suggest a 4-step generation, this often results in degraded output. Setting the generation steps to 10 provides the optimal balance between speed and structural consistency.

1

Download the Turbo LoRA

Locate the MiniMax H3 Turbo LoRA on Hugging Face or Civitai. Download the file and place it in your ComfyUI models/loras directory.

2

Apply the LoRA in Your Workflow

Add the LoRA Loader node to your ComfyUI workflow. Connect it between your model loader and the sampler. Set the strength appropriately (usually around 0.8 to 1.0).

3

Configure Sampler Steps

Despite the LoRA being designed for 4 steps, manually set your KSampler or corresponding video sampler node to 10 steps. Lower steps produce poor, noisy results.

4

Enable Optimizations

Turn Sage Attention ON. Turn Torch Compile ON. Set EFF Sage Attention to OFF (unless your Sigma Shift goes above 0.4).

5

Adjust Sigma Shifts

Set your Sigma Shift values specifically for the media type: 12 for video and 6 for audio. This ensures proper noise scheduling.

MiniMax H3 Workflow Types Explained

MiniMax H3 utilizes a modular workflow system in ComfyUI. Unlike older models where image-to-video and text-to-video share identical pipelines, H3 requires specific models and node configurations based on your input type.

Text-to-Video (T2V)

  • Uses FFLF Model
  • Highly reliable generation
  • Fastest workflow (approx. 4 mins)
  • Excellent prompt adherence

Image-to-Video (I2V)

  • Uses FFLF Model
  • No lip-sync support in this pipeline
  • Requires Torch Compile = True
  • Great for animating static images

Reference-to-Video (Ref2V)

  • Uses Ref2V Model
  • Supports up to 12 inputs (images, audio, video)
  • Best for identity consistency
  • Allows custom audio and lip-sync
The Ref2V Image Discrepancy

When using the Reference to Video workflow, the output will not perfectly match your input image. Unlike LTX or 1.2.2 models that lock the exact pixels of the starting frame, Ref2V generates a close approximation. The skin quality and framing may differ slightly from your reference image.

Workflow Capabilities Comparison

FeatureText-to-VideoImage-to-Video (FFLF)Reference-to-Video (Ref2V)
Model RequiredFFLFFFLFRef2V
Custom Audio InputNoNoYes
Lip SyncNoNoYes
Multiple Image InputsNoNoYes (up to 12)
Exact Frame MatchingN/AYesClose approximation
Background ConsistencyPrompt-dependentHighModerate (use bg reference)

Advanced Settings & Troubleshooting

Fine-tuning the advanced parameters in your ComfyUI workflow is the difference between a 5-minute generation and a 2-hour generation. If your generations are stalling or producing corrupted outputs, the issue almost always lies in your optimization nodes.

Block Cache T8 Optimization

For additional speed, insert a Block Cache T8 node before your Spectrum Apply node. This specifically caches the denoising blocks, significantly speeding up the process. Note that this only speeds up denoising, not the initial encoding phase.

Spectrum Apply Settings

When configuring the Spectrum Apply node, use the following baseline settings for the Turbo LoRA workflow:

ParameterRecommended SettingNotes
Degree1Standard setting for most Turbo generations
Tail Actual Steps1Some users prefer 4; test both for your hardware
Sigma Shift (Video)12Critical for video noise scheduling
Sigma Shift (Audio)6Critical for audio noise scheduling

Troubleshooting Common Errors

If your video quality is poor or the generation fails, follow this diagnostic checklist:

MiniMax H3 Diagnostic Steps:

  • Verify Torch Compile is set to True to prevent Out of Memory errors
  • Ensure you are using GGUF models (Q4) instead of INT8 if experiencing stalling
  • Turn off Spectrum Apply and Block Cache if output quality is degraded
  • Keep Sage Attention enabled even if turning off other optimizations
  • Check that your text encoder is using the Q4 GGUF version for speed

FAQ

Q: Where can I find the MiniMax H3 models for ComfyUI?

You can find all the quantized GGUF models, text encoders, and Turbo LoRAs on Hugging Face. Search for 'MiniMax H3 GGUF' to find Q4, Q5, and Q8 versions of both the First Frame Last Frame model and the Reference to Video model.

Q: Why does my MiniMax H3 generation stall for a long time before reaching the sampler?

This is a common issue when using INT8 or FP8 converted models on consumer GPUs with limited VRAM (like an RTX 3080). Revert to using GGUF quantized models (Q4 or Q5) for both the main model and the text encoder to resolve the stalling.

Q: Can I do lip-syncing with the Image-to-Video workflow in MiniMax H3?

No, the standard Image-to-Video workflow (using the FFLF model) does not support custom audio or lip-sync. To generate videos with custom audio and lip-sync, you must switch to the Reference to Video workflow and load the dedicated Ref2V model.

Q: How many steps should I use with the MiniMax H3 Turbo LoRA?

Although the Turbo LoRA is designed for 4 steps, setting your sampler to 4 steps usually results in poor quality. Set your steps to 10 for the best balance of speed and visual fidelity when using the Turbo LoRA.