MiniMax H3 LoRA Training: A Complete Guide to Fine-Tuning for Custom Video Styles

AI Overview

What is MiniMax H3 LoRA training?

MiniMax H3 LoRA training teaches a small adapter a specific character, visual style, product, or motion while the 33B H3 base stays frozen. The exported adapter is far smaller than a full checkpoint and can be mixed with the base at inference.

How much VRAM do you need to train a MiniMax H3 LoRA locally?

There is no published universal minimum. Treat 16GB as an aggressive reduced-resolution target and 24GB as the more practical starting point; 12GB may work only with heavy quantization, offloading, shorter clips, and substantial system RAM.

How many videos do you need for a MiniMax H3 LoRA dataset?

Cloud trainers accept at least 10 clips; 50–200 clean, varied clips are a stronger practical target. Normalize to 24fps and use valid 17n+5 lengths—22, 39, 56, 73, 90, 107, or 124 frames for the current trainer.

Can you train a MiniMax H3 LoRA without a local GPU?

Yes. Hosted fal trainers accept a ZIP dataset and return a .safetensors adapter. Three published trainer endpoints cover four targets: T2V, I2V/first-frame, first-last-frame inference with an I2V LoRA, and Ref2VA.

Why Train a LoRA for MiniMax H3

MiniMax H3 is a 33B joint video-audio DiT built to generalize across subjects, styles, camera moves, and sound. That breadth is useful, but a general model will not automatically preserve one fictional character, exact product finish, or house motion language across dozens of shots. A LoRA adds a small set of trainable low-rank updates while leaving the base weights frozen.

Train one when repeated prompting and references still drift. Character datasets can teach a stable face, costume, and silhouette across locations. Product datasets can reinforce geometry, materials, logo placement, and preferred ad lighting. A motion LoRA can bias H3 toward a recurring orbit, dance vocabulary, or transition pattern. LoRA is not a repair tool for poor captions or contradictory clips: it amplifies the dataset's patterns, including its mistakes.

MiniMax H3 T2V · official baseline output before custom LoRA training

Use a real baseline render like this to define what the adapter must improve. If the goal is only faster local sampling rather than a new identity or style, use the MiniMax H3 Turbo LoRA guide instead.

The Four H3 LoRA Trainers Explained

H3 has four useful LoRA training targets, but the current hosted API publishes three trainer endpoints. First-last-frame work reuses the I2V adapter family: train with first-frame conditioning, then supply an end frame to the LoRA-enabled I2V inference endpoint.

Choose T2V when the prompt should summon the concept without an image. Choose I2V/FLF2V when the source frame is the identity anchor. Choose Ref2VA when the production brief depends on mixed reference images, video, or audio. For a no-training comparison of the same starting modes, test the Text to Video workspace before committing a dataset.

Dataset Preparation: Clips, Captions, and the 17n+5 Frame Grid

Put videos and captions in one folder with matching stems: shot_001.mp4 and shot_001.txt. Use .mp4, .mov, .avi, or .mkv; keep one concept per clip, remove edits and watermarks, and avoid near-duplicates. Captions should state the persistent concept plus the visible action, camera movement, and audio—not a vague keyword dump. The MiniMax H3 prompting guide provides a useful caption structure.

Normalize every clip to 24fps. H3 uses the frame rule frames = 17n + 5; although 5 is mathematically valid, the current hosted trainer accepts 22–124 frames, so use 22, 39, 56, 73, 90, 107, or 124. Trim first, then verify frame counts rather than relying on duration labels.

Start with at least 10 clips only for a pipeline test. For a useful character or style adapter, 50–200 diverse, well-captioned clips are a better working range. Split out 10–15% for validation. Vary framing, background, camera angle, and motion while keeping the target identity consistent.

A real H3 I2V source from the official workflow. Training data for a product LoRA should preserve details this precisely while varying shot design.

MiniMax H3 I2V · official source-to-motion output

Local Training with ai-toolkit and ComfyUI

ai-toolkit commit 8502a84 added MiniMax H3 T2V and first-frame I2V training. Use that commit or a newer release. It loads the pruned int8 ConvRot H3 transformer and the NVFP4/AWQ Qwen3-VL text encoder, trains with FlowMatch, and exports LoRA keys compatible with ComfyUI. The H3 sampler is guidance-distilled, so leave sampling guidance at 1 rather than tuning conventional CFG.

Use this as a focused starting block inside an ai-toolkit job, not a complete hardware-independent recipe:

network:
  type: lora
  linear: 16
  linear_alpha: 16
train:
  steps: 2000
  noise_scheduler: flowmatch
  optimizer: adamw8bit
  lr: 2e-4
model:
  name_or_path: MiniMaxAI/MiniMax-H3
  arch: minimax_h3
  quantize: true
  model_kwargs:
    partition: fl2va
sample:
  guidance_scale: 1

H3's implementation uses video FlowMatch shift 12 internally and audio shift 3; do not overwrite those with a generic image-model preset. For I2V, enable the dataset's first-frame conditioning and confirm the frame source before a long run. Cache latents/text embeddings, checkpoint every 250–500 steps, and validate the same fixed prompts at each save. ComfyUI is the inspection and inference surface after training; the complete H3 ComfyUI setup covers base files and graphs.

Cloud Training on fal Without a Local GPU

1. Package the paired videos and .txt captions as a ZIP; add the trainer-specific first-frame or reference sidecars when required.
2. Choose T2V, I2V, or Ref2VA trainer according to the table above.
3. Set steps, rank, learning rate, frame count, resolution, and conditioning probability.
4. Run a short test, review validation outputs, then extend only if the concept is still improving.
5. Download the returned .safetensors adapter and config file.

The current schemas default to 2,000 steps, rank 32, learning rate 2e-4, 24fps, and 73 frames; ranks 8–128 are accepted. Billing is calculated per training step, so the UI's live estimate—not a copied blog number—is the reliable cost before launch. The service does not require local VRAM, but dataset quality and checkpoint testing remain your responsibility.

MiniMax H3 Ref2VA · official mixed-reference output

Key Hyperparameters — Rank, Learning Rate, and Steps

Rank is capacity, not a quality slider. Higher rank increases file size and can memorize backgrounds or faces faster. Learning rate controls update size; reduce it when texture becomes harsh or the adapter overwhelms prompts. Steps should be decided by validation, not by reaching a round number. If every prompt reproduces the same composition, roll back to an earlier checkpoint or diversify the data.

Using Your Trained LoRA in Seedance and ComfyUI

For H3 in ComfyUI, copy the adapter to ComfyUI/models/loras/, refresh the model list, load it after the H3 diffusion model, and start at strength 0.7–1.0. Match the adapter to the correct FL2VA or Ref2VA base. Test the trigger phrase against held-out prompts, then compare with the base model at the same seed, frame count, and sampler settings.

Seedance's standard browser workflow does not expose an arbitrary .safetensors loader. The practical bridge is to render a strong LoRA frame or short reference in H3, then use that asset in Seedance Image to Video for browser-based motion, revision, and production. This keeps custom training in the open H3 stack while using Seedance for the faster hosted finish.

Create your next video with Seedance →

Conclusion

MiniMax H3 LoRA training is worthwhile when a repeatable character, product, style, or motion matters more than one good clip: choose the conditioning mode first, build a clean 24fps dataset on the 17n+5 grid, begin with rank 16 and conservative validation, and stop at the best checkpoint rather than the highest step count. Use ai-toolkit for local control or fal when you do not have the GPU, then load the adapter in H3/ComfyUI and pass approved assets into Seedance when you want a simpler browser production path.