MiniMax-H3 Turbo LoRA (larryvrh)
What they did
A LoRA for MiniMax-H3 that renders joint video and stereo audio in as few as 4 steps instead of about 20. The author recommends the v4 (step 600) checkpoint, and notes it can smear fast motion at exactly 4 steps.
Worth knowing
- Price: Open weights on Hugging Face (apache-2.0)
- License: apache-2.0
- Status: sourced. We have not run this ourselves yet (no "Tested" badge).
- Quality bar (computed by code): 5/5 (primary source ✓, reproducible ✓, new ✓, example output ✓, safe ✓)
The best new AI video workflows and prompts, weekly.
Related
▶ demo at sourceHailuo AI
MiniMax's consumer app for its Hailuo video models, with text- and image-to-video plus social-oriented tools.
▶ example at sourceMiniMax video API docs
MiniMax's platform overview, where to find API access for its video, speech and language models.
▶ demo at sourceReplicate video collection
Replicate's curated collection of text-to-video models you can run by API, an alternative host to fal.
▶ demo at sourceHailuo 2.3 Pro on fal
MiniMax Hailuo 2.3 Pro image-to-video on fal at 1080p, billed at a flat price per generation.
▶ YouTube demoLow-VRAM character swap with MiniMax H3 in ComfyUI (no masking)
A ComfyUI workflow that swaps a character in a video using MiniMax H3 plus a Character Swap LoRA, with no masking. It renders at about 0.4–0.5 MP in 8 steps and then upscales to 1.2 MP with the H3 latent upscaler. The creator tested it on an RTX 3060 6 GB with 16 GB RAM: about 20 minutes to generate and 21 to upscale. You build a character sheet first (Qwen Image 2.1, FLUX Krea or H3). Workflow and tutorial are linked from the post.
▶ demo on XRainy anime heartbreak short (MiniMax H3)
A scene-timed vertical anime prompt for a short emotional story, starting with a rain scene at 0–5s.