Minimax H3 R2V With TaoMate - Full Scale Workflow Update - 20260916 (Patreon)

Hey everyone.

If you watched the TaoMate H3 video (https://www.patreon.com/aifuturetech/posts/taomate-h3-3-vs-169715755), you already know the headline: 3 steps, reference-to-video with synced audio.

The free workflow gets you the basic pipeline. This post is the full setup with latent upscale, face tracking, and face stitching — the one I actually used for the final results in the video.

WHAT TO DOWNLOAD (MODEL FILES)

TaoMate LoRA: https://huggingface.co/Kijai/MiniMax-H3_comfy/blob/main/loras/minimax_h3_taomate_3step_lora_avg_rank_19_bf16.safetensors

QUICK START

1. Load the workflow: MinimaxH3_20260916.json

2. Check all loader paths point to the right files (UNETLoader, CLIPLoader, VAELoaders, LoRA loaders)

3. Paste your reference image into the Reference Image input

4. Paste your text prompt into the prompt node

5. Set duration (in seconds) — must be a multiple of 5

6. Set resolution — use the Size Settings Reference note in the workflow for megapixel-to-resolution mapping

7. Set seed (42 is the default, change it for different results)

8. Queue the generation

KEY SETTINGS (AND WHY)

SAMPLER:

- Sampler: euler (standard for MiniMax H3, works best with the LoRA stack)

- Scheduler: simple, 3 steps, 0.45 — this is the TaoMate setting. 3 steps is the minimum the LoRA was trained on. Don't go below 3.

LoRA STRENGTHS:

- TaoMate 3-step: 1.0 (full strength — this is what it was trained for)

- Realism LoRA: 1.0 (enhances face/body detail)

RESOLUTION:

- Default: 0.8 megapixel at 16:9 (1216 x 672)

- For faster generation: drop to 0.5 megapixel (960 x 544)

- For higher quality: go to 1.0 megapixel (1376 x 768)

- All resolutions must be multiples of 32

FPS: 24 (standard, leave as is)

LATENT UPSCALE:

- The paid workflow includes a second pass that upscales from base resolution to 2 megapixels

- This is where the quality jump happens — the base 3-step output is decent but the upscale makes it sharp

FACE REFINEMENT:

- H3FaceTrackCrop: tracks faces across frames using YOLOv8

- H3FaceStitch: re-stitches the refined face back into the video

- Face crop size: 768 x 768 (good balance of detail and speed)

- These nodes are optional — disable them if you want faster generation without face detail

FPS: 24, leave as is.

TIPS

1. Start with the public workflow first. Get the basic TaoMate generation working, then swap to the paid workflow for the upscale and face refinement. Don't try to debug everything at once.

2. The TaoMate LoRA is rank 19 averaged. It's a smaller file (182 MB) than the full rank 128 version. For most use cases it performs just as well. If you want the full rank 128 version, look for the CZMartin22 conversion on HuggingFace (1.24 GB).

3. The AIO Enhancement LoRA at 0.75 is a conservative setting. I tested 0.5 and 1.0 — at 0.5 the enhancement is barely visible, at 1.0 the skin texture gets waxy. 0.75 keeps the detail without over-processing.

4. Duration matters. 5 seconds = 1 segment, 10 seconds = 2 segments, 12 seconds = 2 segments (partial). Each segment uses the same prompt block. For longer videos, use --prompt-json to set different prompts per segment.

5. If you are getting audio desync, check that the audio VAE is loaded as fp32, not fp16. The audio VAE needs full precision for clean sound.

COMMON ISSUES

"Generation hangs after 1 step" — The face refinement nodes are the most likely cause. Disable H3FaceTrackCrop and H3FaceStitch first, get the base generation working, then add them back one at a time.

"Audio sounds distorted" — Your CFG might be too high. TaoMate was distilled at CFG 1.0. Going above 1.0 burns the audio. Drop it back to 1.0.

"Face is blurry after stitch" — The face crop resolution (768 x 768) might be too small for your source resolution. Try increasing the crop size, or check that face_yolov8m.pt is detecting the face correctly.

"Out of memory" — This workflow is VRAM-heavy. 12 GB minimum, 16 GB+ recommended for the full pipeline with face refinement. Drop resolution to 0.5 megapixel if you are tight on VRAM.

Questions? Drop a comment.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji