Troubleshooting

“Model not found” appears

1. Click the status icon at the top right of the window and check “Model in use” and “Load error”.
2. Check that the registered folder is the “MiniMax-H3” folder that directly contains FL2VA (and Ref2VA if you use references). Registering the FL2VA folder itself, or a folder above it, won't load.
3. If you moved the model folder, register its new location with “Add Folder…” in “Model Manager” and select that model.
4. If it fails only when using reference images, videos or audio, Ref2VA may not be downloaded. Add it from Getting the model.

Generation is very slow

Check these in order.

1. Size: larger sizes take longer. Try Small or Medium first.
2. Length: time grows faster than the length (at 512×512 with standard settings, 5 s took about 6 minutes and 15 s about 30 minutes in one measurement).
3. Reference files: generation with reference images, videos and audio tends to take longer, and large reference videos slow the whole generation down a lot.
4. Compute mode: on an M5-class Mac, check that “Fast (int8 cache)” is selected. “Low memory (SSD streaming)” needs less memory but takes longer.
5. Free memory: quit other apps before generating. Running out of memory and swapping slows things down a lot.

“Size” in “Advanced Settings” shows a rough time estimate for the current settings, with a warning when it would take more than an hour.

The download stopped, or there isn't enough space

• The download can resume. If it stopped, click “Resume” in the download window, or reopen the app and start again with the same destination. Files already downloaded are skipped.
• If there isn't enough space, choose a disk with room using “Change…” next to “Save to” (not possible during a download). FL2VA alone needs about 144.1 GB, and about 288.1 GB with Ref2VA. The int8 cache (about 19.3 GB each) is created on the internal disk, separately from the model's location.
• Leaving “Also include Ref2VA” off at first, so only FL2VA is downloaded, keeps the size down.

“Fast mode (experimental)” isn't shown

“Fast mode (experimental)” appears in the advanced settings only on an M5-class GPU (with hardware tensor units) and a build that supports it. The DMG on GitHub is such a build. It is not shown on M4 or earlier chips.

It is a separate feature from the “Fast (int8 cache)” compute mode and the “Fast” speed setting. See the three speed settings.

Where are videos saved?

Why batches need a project

Without an open project, each video is a temporary file that the next generation deletes. A batch would therefore lose all but its last video, so it needs a project to save them in. If you choose 2 or more with no project open, the app asks you to create a new project or pick an existing one. See Projects and batches.

It doesn't run on an older macOS or an Intel Mac

H3cApp is for Apple Silicon Macs with macOS 15 Sequoia or later. It won't launch on macOS 14 or earlier, and Intel Macs are not supported.

Up to v0.4.3 the description said “macOS 13 or later”, but the generation engine actually needed macOS 15, so the requirement was corrected in v0.4.4.

A warning appears at launch

The DMG on the GitHub releases page is signed and notarized by Apple. If you still see a warning, check:

• that you downloaded it from the official releases page;
• the app version (“About H3cApp” in the app menu) and that your Mac runs macOS 15 or later.

We don't recommend turning off Gatekeeper or similar. If unsure, see Apple's guide Safely open apps on your Mac.

Updating from an earlier version

Open the new DMG and drag H3cApp into Applications to replace the old one. Settings, presets and registered models carry over. Up to v0.3.0 the app was called “h3c-app”, so once you have v0.4.0 or later installed you can delete the old h3c-app.app. In v0.6.0 the folder for models and caches is renamed too, from ~/Library/Application Support/h3c-app to H3cApp. The app renames it on its first launch and updates the locations of registered models and LoRAs, so there is nothing to move by hand.

Reporting a problem

Please use GitHub Issues. This information helps:

• App version (e.g. 0.4.7), macOS version, chip (e.g. M5), memory
• How you generated (from text or images), shape, size, length, and advanced settings (compute mode, speed, fast mode, …)
• What you did and what happened (steps to reproduce), and the exact error message

Issues are public. Leave out prompts you don't want to share and file locations that include personal names.

Licenses and terms

• The app (H3cApp): MIT License, based on h3.c by Salvatore Sanfilippo (antirez). Notices for bundled software are in THIRD_PARTY_NOTICES.md.
• The video model (MiniMax H3): H3cApp generates videos with MiniMax H3 by MiniMax. The model comes under its own, somewhat unusual license, the MiniMax H3 Community License, separate from the app, and it differs from typical open-source licenses. The main points (a summary):
  • The license defines the regions it applies to and excludes the European Union, the United Kingdom, the Republic of Korea and the United States.
  • An Acceptable Use Policy lists prohibited uses.
  • The model and its outputs may not be used to improve other AI models.
  • Commercial products and services have extra conditions (prior permission above a revenue threshold, showing “MiniMax H3” in the interface, and more).
  Before use, read the full license and the provider's license Q&A.
• Add-on models (LoRA): follow the terms of each one's source.

This site can't tell you whether commercial or other uses are allowed; check the latest licenses above. The app is distributed free of charge, but the data for downloading the model and the electricity for generating are on you.