Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.
LoRAs
A LoRA is a small add-on that teaches a model a face, a style, a motion or a speed-up. The LoRAs section of the Video form (it moves into Images too) is where you turn them on.
The picker {#picker}
- The summary reads none active, or how many are installed and active.
- Click a row to turn a LoRA on or off. Turning one on adds its first trigger word to your prompt; the trigger chips on an active row add them again.
- strength — a slider from -2 to 2. It starts at the LoRA's recommended strength, or 1.0.
- With five or more LoRAs, a filter box finds them by name or trigger word.
- The row buttons rename it (display name only), download the file, open it on CivitAI, or delete it from disk (asks first — it is permanent).
- Write a guide has the planner model write what the LoRA does, how to prompt it, and a strength to start from.
- Enhance knows the LoRAs that are on: it reads each one's guide (or the creator's description from CivitAI / Hugging Face) as reference, keeps their trigger words, and puts a trigger word back if the rewrite left it out. The line under the result says which.
- ? means its family is unknown — it may or may not work on this engine. Update means CivitAI has a newer version (Check for updates asks).
LoRAs are per engine: an LTX LoRA cannot load on Hailuo H3 and an H3 LoRA cannot load on LTX. The picker shows the active engine's library and offers Show N from other modes.
Browse CivitAI and Hugging Face {#browse}
Browse CivitAI opens the browser.
- Pick the source — CivitAI or Hugging Face — and the engine: All, LTX or Hailuo H3.
- Search (a name, a style, a creator; on Hugging Face also
author:someoneorowner/repo) and press Enter. - Install on a result. It lands in
mlx_models/loras/and turns on in the picker — unless it is for the other engine, in which case it says which engine to switch to.
CivitAI needs an API key to download: paste it in the browser's banner or in Settings → API tokens. Hugging Face results can be filtered by kind (Characters, Styles, Motion, Speed) and With example; read the repo before you install.
Adding a file yourself {#import}
- LTX — copy
.safetensorsfiles intomlx_models/loras/and press the rescan button. - Hailuo H3 — Import H3 LoRA takes a
.safetensorsfile (up to 4 GB) and reports how it was read and its recommended strength. H3 LoRAs live in their own library,mlx_models/hailuo-h3/loras/; you can also drop files there by hand and press rescan, with the same checks. H3 accepts the runner layout with pairedlora_A/lora_Btensors. ComfyUI repacks and Kohya files (lora_down,lora_upand.alpha) are converted on import, and a Kohya file keeps its alpha scaling, so strength 1.0 means "as trained". H3 LoRAs from CivitAI, Hugging Face, fal, or your own imports store one internal set of weights in a different order than H3 expects; Phosphene checks each file and, when needed, runs a reordered copy kept in arunner-layoutfolder next to the library file, leaving your file untouched. This brings character LoRAs and Fast speed mode closer to how they were trained, most visibly with sharper mouths.
Stacking LoRAs on Hailuo H3 {#h3-stacking}
- On an H3 install that stacks, up to 4 LoRAs per render, each with its own strength, and Turbo rides along. Keep the strengths' total near 1.5 or under, and avoid two LoRAs that pull the same thing (two faces, two styles).
- On an older H3 install there is one adapter slot: the Adapter row chooses Turbo or My LoRA for this render. Picking more than the install allows stops the job with a message saying how many to un-pick.