Testing MiniMax H3: Generating High-Consistency Yu-Gi-Oh Anime with 9 Reference Images
circlenline · reddit · 2026-08-05
Using the MiniMax H3 video generation model (ref2v mode) locally on an RTX 5080, a developer successfully generated a 14-second Yu-Gi-Oh themed animation using only 9 reference images.
The test demonstrated exceptional multi-subject consistency:
- Characters & Scenes: Perfectly preserved the custom protagonist's outfit, Kaiba's arrogant posture and classic expressions, and the formation of the three Blue-Eyes White Dragons.
- Complex Choreography: Utilized highly detailed structured prompts to define storyboard compositions, camera switching, and the narrative (protagonist summoning the God of Nodes to defeat Kaiba).
- Object Replication: The model accurately "replicated" the 5 physical card illustrations from the reference image without redesigning them.
The author also shared specific local deployment configs, including FP8 pruned diffusion models, Qwen3VL text encoder, and sampler parameters.
More from Multimodal
- MiniMax Video Model Test: Generates 8 Coherent Clips from a Single Prompt — intermundia · 2026-08-05
- Minimax H3 Video Generation Tested: Usable on First Gen, Beating Wan2.2 — R34vspec · 2026-08-05
- Testing MiniMax H3: Generating High-Quality Arabic Motion Graphics in One Go — aziz4ai · 2026-08-05
- MIRA: A Fully AI-Generated Rocket League Game Playable in Browser — mathemagic1an · 2026-08-05
- Analyzing the Four Core Paradigms of Modern In-Context TTS — rdesh26 · 2026-08-05
- FLUX 3 Video Tested: Generates 20-Second Cinematic Animation from a Single Prompt — aziz4ai · 2026-08-05