MiniMax H3 FHD / 2.5K Detailer — ComfyUI Workflow >>> zuanfilm_H3_HD_2K_detailer_v4.json

A high-detail MiniMax H3 reference-to-video workflow for ComfyUI for generating and refining FHD/2.5K high-resolution H3 video while preserving reference-image information, motion, and native audio.

The workflow combines MiniMax H3, reference-to-video conditioning, a low-step RES4LYF sampling setup, H3-specific attention optimizations, latent upscaling, and a final image-detail / tone-processing stage.

Check a sample video: https://www.youtube.com/watch?v=epjELHgEH_o

✨ What is this?

This workflow is designed for FHD/2.5K high-detail MiniMax H3 video generation and refinement in ComfyUI. Basically you have 3 workflows in one: low-res FL2VA, low-res Ref2VA and FHD-2.5K-Detailer, to use any of them you bypass the other 2 groups

The workflow supports reference-image driven generation, allowing multiple visual references to be supplied to the H3 pipeline, so you can create a low res video (0,5 or 0.6Mp) and use that generation as base to create your Full HD video, you can also add for example face references so the detailer will make face-swaps

It was built with an emphasis on:

• High visual detail

• Reference-image consistency

• Short-step H3 generation

• Native audio preservation

• Lipsync/audio locking

• Memory-efficient attention

• Latent-space upscaling

• Final texture and tone refinement >>> Orion4D

• Practical GPU memory management

---

🎬 Workflow Overview

The general pipeline is:

Reference Images
       │
       ▼
MiniMax H3 Reference-to-Video
       │
       ▼
H3 Model + LoRA
       │
       ▼
Attention / Memory Optimizations
       │
       ▼
H3 Flow Shift
       │
       ▼
Switch  AUDIO (Impact)
       │
       ▼
H3 Latent Upscaling
       │
       ▼
RES4LYF 4-Step Sampling
       │
       ▼
VAE Decode
       │
       ▼
Texture / Tone Processing >>

Orion4D · DCTL Tone Mapper APP
Node #702 — JSON presets stored in /custom_nodes/Orion4D_FXMax/presets/color_fx_dctl_tone_mapper
Zuanfilm-Reinhard H3 Detailer.json

Orion4D · Texture APP
Node #703 — JSON presets stored in /custom_nodes/Orion4D_FXMax/presets/Sharpness_Texture
Minimax-H3-27.json

Orion4D · SharpnessMax APP
Node #704 — JSON presets stored in /custom_nodes/Orion4D_FXMax/presets/Sharpness_SharpnessMax
Soft Portrait.json


       │
       ▼
Video Combine
       │
       ▼
Final MP4

---

🚀 Main Features

MiniMax H3 Reference-to-Video

The workflow uses the MiniMaxH3ReferenceToVideo node as the main reference-to-video conditioning stage.

The included configuration uses:

• Width: 1344
• Height: 768
• Length: 124 frames
• Reference image size: max

The workflow also includes a text prompt input for describing the desired motion and performance.

---

🖼️ Multiple Reference Images

The workflow includes multiple image inputs that can be used as visual references.

The current workflow contains dedicated inputs labeled:

• H3-FIRST FRAME / IMAGE 1
• H3-LAST FRAME / IMAGE 2
• H3 IMAGE 3
• H3 IMAGE 4
• H3 IMAGE 5

The supplied example references are loaded at 1344 and aligned to a 32-pixel multiple.

You can replace these images with your own references, if you generate the base low res video on the FL2VA workflow, you can use the exact same references ans prompt in the HD detailer.

---

⚡ 4-Step RES4LYF Sampling

The workflow uses:

ClownsharKSampler

with the following configuration:

Sampler:    multistep/res_2m
Scheduler:  beta57
Steps:      4
Denoise:    0.45
CFG:        1
Eta:        0
BongMath:   enabled

The workflow also includes a SigmasSchedulePreview so the sampling schedule can be inspected.

This low-step configuration is intended to make H3 generation substantially more practical while retaining the characteristic H3 motion and visual behavior. I tested all samplers and res_2m is the most precise keeping memory and time efficiency, if you want a faster generation with a little less detail you can use 3 steps instead 4 er_sde/beta 57 is also a good combination but will lose some detail and even will affect character acting, audio and motion consistency

---

🧠 Memory & Attention Optimization

The workflow contains several optimizations intended to make MiniMax H3 more practical on constrained GPUs.

Included components include:

MiniMax Low VRAM Attention

MiniMaxLowVRAMAttention
head_chunks = 4

Chunked Feed Forward

MiniMaxChunkFeedForward
chunks = 4
seq_threshold = 4096

Memory-Efficient Sage Attention

MiniMaxH3MemoryEfficientSageAttentionPatch
I bypass this node but some people use this so is include

H3 Sparse Local Attention

H3SLAAttention
sparsity_ratio = 0.9
block_size = 64
min_seq_len = 8192
protect_audio = true
enabled = true

Sparce Local attention will reduce a lot the generation time but obviously will affect quality so you can bypass this node if want Top HD quality The workflow also uses explicit RAM and VRAM cleanup nodes.

---

🔧 H3 Model Configuration

The workflow uses the MiniMax H3 INT8 ConvRot diffusion model:

minimax_h3_ref2va_pruned_int8_convrot.safetensors

and the Qwen3-VL 32B MiniMax H3 text/vision encoder:

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

The workflow uses the MiniMax H3 video VAE:

minimax_h3_video_vae_int8_convrot.safetensors

and the MiniMax H3 audio VAE:

minimax_h3_audio_vae_fp32.safetensors

Model files should be obtained from their respective official/model-author sources. This repository does not redistribute those model weights.

---

🧩 H3 LoRA

The workflow contains an H3 model-only LoRA pass:

minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors

you also can test this one

minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors

with:

strength_model = 0.5

This is used as part of the low-step H3 generation configuration.

---

📐 Resolution

The workflow contains a resolution selector and an internal reference table for 16:9 output.

All dimensions are aligned to a 32-pixel multiple.

The current workflow uses approximately 0.8 MP / 1344 × 768 for the H3 working resolution and 2.1 MP / HD for the latent upscaling stage.

---

🔬 H3 Latent Upscaling

The workflow includes:

MinimaxH3LatentUpscaler3D

configured for:

Model:
minimax_h3_latent_upscaler_3d_bf16.safetensors

Mode:
megapixels

Target:
2.1 MP

Alignment:
32

Device:
CUDA

Precision:
BF16

This is the main stage responsible for moving the H3 result toward the HD output resolution while remaining in latent space.

---

🎨 Final Detail & Image Processing

After the H3 latent is decoded, the workflow applies a final image-processing stage using Orion4D FXMax.

This stage is used to refine the tonal response, contrast, highlight behavior, shadow detail, saturation, and perceived image quality of the generated frames.

DCTL Tone Mapper

The workflow includes a custom preset:

Preset: zuanfilm-H3 Mode: Reinhard

The current preset uses:

Setting Value Exposure 0.09 Contrast 0.81 Pivot 0.69 Highlight rolloff 0.27 Shadow lift 0.33 Black floor 0.12 Saturation 0.93 Strength 0.30

The purpose of this stage is not to fundamentally change the generated image, but to provide a controlled finishing pass after H3 generation and decoding.

Texture Processing

A second Orion4D FXMax stage applies the workflow's custom texture treatment:

Preset: Minimax-H3-27 Texture strength: 0.27 Texture radius: 0.30

Additional clarity, sharpening, noise reduction, and high-pass parameters are included in the preset.

The texture stage is designed to complement the H3 output rather than replace the detail generated by the model.

Processing Pipeline

The complete finishing section is:

H3 Latent Upscaling ↓ VAE Decode ↓ DCTL Tone Mapper Preset: zuanfilm-SeedVR2-H3 ↓ Texture Preset: Minimax-H3-27 ↓ Video Assembly

These presets are part of the workflow's visual finishing approach and can be modified directly through the Orion4D FXMax controls if a different look is desired.

These stages are intended to refine perceived texture, clarity, contrast, and overall image quality after H3 generation.

They are not part of the MiniMax H3 model itself.

---

## 🎥 Output

The workflow combines the processed frames using:

```text
VHS_VideoCombine

Current example configuration:

Format:       H.264 MP4
Frame rate:   24 FPS
Pixel format: yuv420p
CRF:          15
Save metadata: enabled

Audio is passed into the final video combine stage.

---

📝 Prompt

The workflow includes a multiline prompt input. I recommend you to use a robust prompt enhancer to replace the Context IR used by MINIMAX H3 API, I designed a Context Assist using 37000 tokens of filmmaking reasoning:

https://chatgpt.com/g/g-6a72ee44de7481919daeee5879b328cc-zh3-gpt

The example prompt is:

a woman from image 1 is walking with her dog and then the walk to the camera with intense expression. soft lateral light show their face and outfits

Replace this with your own scene description and motion direction.

For best results, describe:

1. The subject
2. The action
3. The movement
4. The interaction with the camera
5. Facial expression / performance
6. Lighting
7. Clothing or appearance when important

---

🛠️ Required Custom Nodes

This workflow is not a ComfyUI-only workflow. Several custom node packages are required.

The workflow uses nodes from:

• RES4LYF
• ComfyUI-FearnworksNodes
• ComfyUI-KJNodes
• plaguekind-nodes
• Orion4D FXMax
• VideoHelperSuite
• ComfyUI core nodes

Some optimization nodes are specifically associated with MiniMax H3.

Install the required custom nodes before loading the workflow. Exact node availability and compatibility may change as ComfyUI and the custom node repositories evolve.

---

📦 Models

The workflow expects the following model components.

Diffusion Model

minimax_h3_ref2va_pruned_int8_convrot.safetensors

Text / Vision Encoder

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

Video VAE

minimax_h3_video_vae_int8_convrot.safetensors

Audio VAE

minimax_h3_audio_vae_fp32.safetensors

H3 LoRA

minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors

H3 Latent Upscaler

minimax_h3_latent_upscaler_3d_bf16.safetensors

Preview VAE

taeh3.safetensors

Model files are not included with this workflow.

---

💻 Recommended Starting Settings

For a first test, use the workflow close to its included configuration:

Working resolution: 1344 × 768
Frames:             124
Sampler:            multistep/res_2m
Scheduler:          beta57
Steps:              4
Denoise:            0.45
CFG:                1
Flow shift video:   12
Flow shift audio:   3
Latent upscale:     2.1 MP
Alignment:          32
Precision:          BF16

Start with the included settings before changing multiple parameters simultaneously.

---

⚠️ GPU Memory

MiniMax H3 is a large model and the workflow contains several memory-optimization stages specifically because of this.

Your actual VRAM requirement will depend on:

• Input resolution
• Number of frames
• Reference images
• Attention implementation
• Precision
• Model loading configuration
• ComfyUI version
• Installed custom nodes
• GPU architecture

The workflow should therefore be considered GPU-intensive.

If you encounter out-of-memory errors, reduce the working resolution and/or frame count before increasing the final output resolution.

---

🧪 Workflow Philosophy

The goal of this workflow is not simply to generate a video at a larger resolution.

The pipeline separates the process into:

H3 Generation
      ↓
Latent Refinement / Upscaling
      ↓
VAE Decode
      ↓
Image Detail Processing
      ↓
Video Assembly

This allows the expensive H3 generation stage to operate at a more manageable working resolution while the final detailing process happens later.

---

📌 Notes

• This is a community workflow, not an official MiniMax workflow.
• Model weights are not redistributed here.
• Custom node compatibility may change over time.
• Your ComfyUI installation should be reasonably up to date.
• Some nodes in the workflow are specifically optimized for MiniMax H3.
• The workflow contains example images and example prompts; replace them with your own assets.
• The example output settings are optimized for the author's workflow and may require adjustment for different GPUs.

---

🙏 Credits

This workflow builds on the work of the developers and communities behind:

• MiniMax H3
• ComfyUI
• RES4LYF
• ComfyUI-KJNodes
• VideoHelperSuite
• Fearnworks nodes
• plaguekind nodes
• Orion4D FXMax

Please support the original developers and model authors.

---

👤 Author

Zuan Carreño — @zuanfilm

⭐ If this workflow helps you

If you find this workflow useful, consider:

• ⭐ starring the repository
• sharing your results
• reporting reproducible issues
• crediting the workflow when publishing derivatives

Have fun creating.