A MiniMax H3 LoRA run can fail after you have paid for it, for reasons that have nothing to do with your idea: a zip with one photo in it, clips that are too short to sample, or an adapter trained for the wrong generation mode. This guide covers the decisions to make before you upload: which of the four trainers to use, how to build the clip set, how many steps to run and how to use the adapter afterwards.
Quick answer: MiniMax H3 has four LoRA trainers (text-to-video, image-to-video, first-and-last-frame and reference-to-video), and an adapter only loads in the mode it was trained for. Train on 10 to 50 video clips only (no stills), each at least about 3 seconds long, with 1,000 to 2,000 steps for a first run. On Imagera's LoRA Trainer all four run in the browser, with the credit cost shown before you start.
What is MiniMax H3?
MiniMax H3 is a video-and-audio model. MiniMax's open-source announcement (3 August 2026) lists 4 to 15-second clips at 24 fps, a 768p baseline with output up to 2K, 32 kHz stereo audio, and a reference mode that takes up to 9 images, 3 video clips and 3 audio clips, 12 files at most (MiniMax H3 open-source release). The open weights come under the MiniMax H3 Community License Agreement, which carries its own usage restrictions. Read it before you plan a local pipeline.
Which MiniMax H3 trainer should you use?
Decide how you will generate first; the adapter is bound to that mode.
| Trainer on Imagera | Learns | Use it when | Steps (min–max, default) | Typical run |
|---|---|---|---|---|
| MiniMax H3 Text→Video | A style, subject or motion you summon from a prompt | No input image at generation time | 1,000–4,000, default 2,000 | ~40 min |
| MiniMax H3 Image→Video | Motion applied to a supplied still | You animate product shots, portraits or artwork | 1,000–2,000, default 2,000 | ~45 min |
| MiniMax H3 First+Last Frame | The journey between two stills | You need a defined start and end frame | 1,000–2,000, default 2,000 | ~45 min |
| MiniMax H3 Reference→Video | How to place a subject into new shots from references | You compose shots from up to 9 reference images | Fixed at 1,000 | ~50 min |
If you are unsure, start with Text→Video: it is the most flexible and tells you quickly whether your dataset carries the look.
How to build a MiniMax H3 LoRA dataset
The rules below come from the trainer's own configuration; the first three cause most failed runs.
- Video only. A zip with images in it is rejected, and so is a mix of images and clips. If you only have photos, use reference mode instead of training.
- Clips long enough to sample. The trainer cuts 73 frames from each clip at 24 fps, so any clip shorter than about 3 seconds cannot be used. Conform footage to 24 fps and keep each clip at 3 seconds or more.
- 10 to 50 clips. Ten is the minimum; 15–30 is the recommended range.
- One continuous shot per clip. No scene cuts mid-clip, no watermarks.
- Captions are optional. Put
clip01.txtnext toclip01.mp4and it is matched by name. Describe subject, action, setting, lighting and camera in plain sentences. - Mode-specific framing. For Image→Video, favour clips whose first frame is sharp and uncluttered. For First+Last Frame, both ends of each clip matter. For Reference→Video, show the subject from several angles and distances.
Illustration: a character LoRA's job is to keep one subject consistent across every frame and every clip.
How many steps should a video LoRA run?
- First run: the default (2,000 on Text→Video, Image→Video and First+Last Frame; Reference→Video is fixed at 1,000). You are checking that the dataset teaches the right thing.
- Adjust only when you can name the gap. More steps sharpen what the clips contain; they cannot add what the clips lack. If the style is missing, add or re-cut clips before adding steps.
Character LoRA or reference images?
Reference mode needs no training: supply up to 9 images and the model composes from them. It suits a handful of shots of one face or product. Train a character LoRA when you need the same identity across many clips, when references drift from shot to shot, or when you want a movement or camera behaviour that a still image cannot show.
Using the adapter
Generate in LoRA Generate, choosing the same MiniMax H3 mode you trained. Up to three adapters load in one generation, for example a style adapter with a character adapter. Start each around 0.6–0.8 strength; two adapters at full strength tend to fight.
Train locally instead? AI Toolkit lists MiniMax H3 among its video models; plan for a high-memory GPU and check the licence terms first.
What does a MiniMax H3 LoRA cost?
The trainer quotes the credit cost before each run; it scales with the trainer and the step count, and Reference→Video is the heaviest of the four. Budget for two runs: the first usually shows what the dataset is really teaching. Plans start at $19.99/mo on the Starter plan and credit packs start at $39.99 (pricing).
Train a MiniMax H3 LoRA in the browser →
For fundamentals that apply to every model, see how to train a LoRA online; for the model itself, MiniMax H3 explained.
