Summary
This video from the Veteran AI channel puts three community-made Turbo LoRAs for the MiniMax H3 video generation model head-to-head: Kijai’s widely-used ComfyUI-optimized version, larryvrh’s V4-600 EMA (officially recommended by MiniMax on X), and drbaph’s variant specifically tuned for the pruned INT8 model. All tests run under identical conditions inside a ComfyUI workflow on RunningHub, using the same prompt, seed, and scene to isolate quality differences.
The benchmarks reveal a dramatic speed gain — generation time on an RTX 4090 drops from roughly 8 minutes 29 seconds to just 1 minute 38 seconds using four-step Turbo sampling. While acceleration is roughly equal across all three LoRAs, quality diverges significantly: Kijai’s version produces electronic noise in audio, larryvrh’s is cleaner but still shows voice-over drift without lip-sync, and drbaph’s model delivers the most convincing lip movement and cleanest audio. The creator recommends drbaph as the default choice for gentle-motion scenes.
The second half of the video addresses a common failure mode — flickering during fast or dramatic action sequences — and introduces a two-stage sampling workflow as the fix. This technique splits the diffusion process into two distinct phases to stabilize motion, making it a practical companion to the LoRA comparison for anyone pushing MiniMax H3 beyond simple scenes.
📺 Source: Veteran AI · Published August 11, 2026
🏷️ Format: Comparison
