ComfyUI MiniMax H3 Video Refinement Workflow - 20260912 - For Patreon Supporters (Patreon)

Hey everyone. The public video covers what H3 Face Refine does and the results.

https://www.patreon.com/aifuturetech/posts/fix-blurry-faces-169324729

This post gives you the workflow file, the model paths, settings, and tips. Also have the character sheet attached, in-case you miss previous tutorial how to create:

----------------------------------------

RESOURCES

----------------------------------------

MiniMax H3 Video Prompt Agent Skill

https://github.com/benjiyaya/Minimax-H3-Prompt-AgentSkill

Calliope - Local-first AI Idea-to-video studio

https://github.com/benjiyaya/Calliope

ComfyUI-H3-FaceRefine (the custom node pack):

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

Ultralytics YOLO models (face detection):

https://huggingface.co/Ultralytics/models

Face detector used in this workflow:

https://huggingface.co/Bingsu/adetailer

MiniMax H3 models:

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

Turbo LoRA:

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

Realism LoRA:

https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA

----------------------------------------

WHAT TO DOWNLOAD

----------------------------------------

Face Detection:

Download: face_yolov8m.pt

Place in: models/ultralytics/bbox/

Source: https://huggingface.co/Bingsu/adetailer

MiniMax H3:

models/diffusion_models/minimax-h3/minimax_h3_ref2va_int8_convrot.safetensors

models/text_encoders/minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors

models/vae/minimax-h3/minimax_h3_video_vae_fp16.safetensors

models/vae/minimax-h3/minimax_h3_audio_vae_fp32.safetensors

Turbo LoRA:

models/loras/minimax-h3/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

Strength: 1.0

Realism LoRA:

Download: h3-realism-people-t2v-i2v-r2v.safetensors

Place in: models/loras/minimax-h3/

Strength: 0.95

Note: The workflow references this as h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors (local rename). Rename your download or update the loader path.

Frame Interpolation:

models/upscale_models/film_net_fp16.safetensors

Custom Nodes Required:

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

ComfyUI with native MiniMax H3 nodes

ComfyUI-VideoHelperSuite

ComfyUI-KJNodes

ComfyUI-Impact-Pack

ComfyUI-Frame-Interpolation

ComfyUI-Easy-Use

DLSS5 custom node

rgthree

----------------------------------------

QUICK START

----------------------------------------

1. Download all model files listed above.

2. Drag the JSON into ComfyUI.

3. Check every loader node — confirm model paths match your folder structure.

4. Load reference images into the VHS_LoadImagePath nodes.

5. Write your text prompt in the PrimitiveStringMultiline node.

6. Set Resolution Selector (default: 16:9 at 0.8 megapixels).

7. Set Duration (default: 12 seconds).

8. Click Queue.

----------------------------------------

PIPELINE STAGES

----------------------------------------

The workflow runs in this order:

1. User Inputs — reference images, prompt, duration, resolution, audio

2. LoRAs — Turbo (1.0) + Realism (0.95)

3. Patches — SageAttention, ChunkFeedForward

4. Conditioning — prompt encoding

5. 1st Sampling — H3 ReferenceToVideo (simple, 8 steps, denoise 1.0)

6. Decode 1st pass — VAEDecode, VAEDecodeAudio

7. 2nd Stage — H3PerFrameDenoise, H3InjectVideoLatent (denoise 0.45)

8. Face Track — H3FaceTrackCrop (face_yolov8m.pt, confidence 0.3, crop 2.5, canvas 768)

9. Face Refine Sampling — second pass on tracked face crops

10. Decode refined — VAEDecode of refined faces

11. Stitch back — H3FaceStitch (face_only, feather 24, colour_match 1.0)

12. Frame Interpolation + Upscale — FILM (2x) + DLSS5 (1.5x Quality)

----------------------------------------

KEY SETTINGS

----------------------------------------

H3FaceTrackCrop:

detector: face_yolov8m.pt

confidence: 0.3

crop_factor: 2.5

canvas: 768x768

smooth_window: 21, size_smooth_window: 51

select: largest_face

H3PerFrameDenoise:

small_face: 1.0, large_face: 0.35

face_px_small: 30, face_px_large: 120

smooth_frames: 9

IMPORTANT: takes MODEL in, returns MODEL out. Patched model must reach BasicGuider.

BasicScheduler:

face refinement: simple, 8 steps, denoise 0.45

first pass: simple, 8 steps, denoise 1.0

H3FaceStitch:

paste_region: face_only, feather: 24, colour_match: 1.0

LoRAs:

Turbo 1.0 — speeds to 4-8 steps

Realism 0.95 — skin texture. Drop to 0.8 if over-sharpened.

Duration: 12 seconds default. Frame count auto-snapped to H3 grid.

----------------------------------------

TIPS

----------------------------------------

1. Turbo at 1.0 with 8 steps. Above 8 steps wastes time.

2. Denoise 0.45 is the sweet spot. Below 0.35 too gentle, above 0.45 faces drift.

3. Crop factor 2.5 — seam shows below 2.0, magnification lost above 3.0.

4. Cut detection off by default. Set to auto (pyscenedetect) for hard cuts.

5. Cache clear nodes between stages prevent OOM. Do not delete them.

6. FILM doubles frame count. A 12s clip becomes 24s at 48fps.

----------------------------------------

COMMON ISSUES

----------------------------------------

Faces still blurry — H3PerFrameDenoise must be in model path to BasicGuider.

Crop shivers — raise smooth_window to 31 or 41.

Seam visible — feather 24 for rect. Raise to 30 if hard edge.

Face drifts — drop denoise from 0.45 to 0.35.

OOM — drop canvas to 512. Cache clears are essential.

DLSS5 not found — needs NVIDIA RTX GPU.

----------------------------------------

Questions? Drop them in the comments with your GPU, VRAM, and the exact error.

Thanks for supporting the channel.

— Benji