Local Fine-Tuning MiniMax H3 Video Model Tested on a Single RTX 5090
No_Statement_7481 · reddit · 2026-08-07
A developer shared their practice of locally fine-tuning the MiniMax H3 video model using the Osris Ai toolkit. They built a dataset of 1-second, 512x512 resolution videos for LoRA training tests.
Hardware & Performance:
- With Layer offloading enabled on a single RTX 5090, VRAM usage was extremely low (6.6GB / 32GB VRAM), but consumed 78GB of system RAM.
- Each iteration step took about 3.43 seconds, meaning 3000 steps will take roughly 3 hours.
Pitfalls & Tuning:
- Frame count is crucial: failing to adjust the default 39 frames to the dataset's 22 frames resulted in samples with "chipmunk sounding" audio.
- The author currently used simple captions and plans to test the limits with a better dataset and proper prompting standards next.
Related event: MiniMax H3 Video Model Supports Local LoRA Training on 16GB VRAM(3 posts)→
More from Multimodal
- Kling 4.0 fully available, shows markedly better motion coherence in fast-action scenes — alifcoder · 2026-10-16
- MJev-Spatial: Single-Pass Classification Beats Autoregressive VLMs on Spatial Reasoning — mindlab-research · 2026-10-16
- H-DDCM: Hadamard Codebooks Make Diffusion-Based Image and Video Compression Scalable — Technion · 2026-10-16
- ComfyUI v0.41.0 ships with Flux3 image model and MiniMax-H3 LightVAE support — Fcking_Chuck · 2026-10-16
- AI-generated short film 'The Dragon Rider' glides cinematically above the clouds — jijo6686 · 2026-10-16
- Realistic candid Midjourney portrait prompt: natural skin, lived-in bedroom light — tisch_eins · 2026-10-16