Fine-Tune MiniMax H3 on Your Brand Data

Why customize

Where general models fail in production

Workload	What must stay consistent
Performance ads	Products, logos, packaging and brand colors across many variants
Game marketing	Characters, art style and gameplay look across creative tests
Short drama	Characters, wardrobe and setting across shots and episodes

Each rejected clip costs a full generation and a review. Consistency problems are the most common reason for a reject, so they drive cost.

How we customize

Three layers of customization

01

References at generation time

FastH3-Omni, the open-weight, accelerated MiniMax H3 with full omni reference input, takes text, up to 9 reference images, up to 3 reference videos, up to 3 audio references and first and last frames in one generation. Nuva Lab built it with FastVideo and runs the production version, with dedicated deployment, fine-tuning on your data and the Creative Agent end to end. FastH3-Omni preview.

02

Post-training on your data

We fine-tune or post-train the model on your approved clips and assets, so your look becomes the model's default instead of a long prompt.

03

Evaluation on your criteria

We build a test set from your own review decisions and check every model update against it before it reaches production.

Choose the method

LoRA training vs post-training vs references

Method	What changes	Best for
LoRA adapter	A small set of added weights; the base model stays the same	One character, product or style
Full post-training	The model weights	A house style across many workloads
FastH3-Omni references	No weights; references go in at generation time	Per-job consistency and fast changes

Most teams start with references and LoRA adapters, then move to post-training when one style covers many workloads. We choose the method from your review data, not from a fixed plan.

Ownership

Your data, your model, your deployment

Your assets, your customized weights and your review data stay in your dedicated deployment. The customized model runs on a host that only you use.

Customization follows the base model license. For MiniMax H3, fine-tunes and LoRAs are Model Derivatives, so the H3 territory rule applies. Nuva Lab holds a MiniMax authorization for H3 deployments that serve US companies.

FAQ

Questions

Can I fine-tune MiniMax H3 on my own data?

Yes. H3 is open weight, so it can be fine-tuned or post-trained. The result is a Model Derivative under the MiniMax H3 license, so the license territory rule and restrictions still apply.

How do you train a LoRA for MiniMax H3?

Collect approved clips and images of one subject or style, caption them, train a LoRA adapter on the open H3 weights, and test it on held-out briefs. Nuva Lab runs this process on your dedicated deployment.

Is training a LoRA expensive?

A LoRA trains only a small set of added weights, so it needs much less compute than full post-training. Data preparation and evaluation are often the larger cost. The total depends on clip count, resolution and length.

What is FastH3-Omni?

FastH3-Omni, the open-weight, accelerated MiniMax H3 with full omni reference input, takes text, up to 9 reference images, up to 3 reference videos, up to 3 audio references and first and last frames in one generation. Nuva Lab built it with FastVideo and runs the production version, with dedicated deployment, fine-tuning on your data and the Creative Agent end to end. FastH3-Omni preview.

How do you measure whether customization works?

We measure the acceptance rate on a held-out set of your own briefs, scored with your review criteria, before and after each change.

Sources and further reading

Fine-tune MiniMax H3 on your brand: LoRA, post-training and references