Testing MiniMax H3's Celebrity Character Generation on RTX 5090

JuniorEnsign · reddit · 2026-08-11

A developer tested the MiniMax H3 model's ability to generate realistic celebrity characters for 5-second, 1MP text-to-video clips. The test was conducted locally using ComfyUI with INT8 quantization on an RTX 5090.

Test Details:

• The prompt was relatively simple, setting up a TV interview scene with a static camera and dark blue background, requiring the specific celebrity to say an aggressive line.
• The methodology emphasized "one run per person, no picking the best results" to reflect the model's raw baseline performance.
• The model notably failed to generate actor Christoph Waltz, who was subsequently excluded from the final test list.

Original post →

More from Multimodal

• SIGGRAPH 2025 Paper: Turning Motion-Blurred Photos into Coherent Videos — jon_barron · 2026-08-11
• Real-Time Mocap Breakthrough: Capturing Bodies, Fingers, Faces, and Multi-Character — andrew_n_carr · 2026-08-11
• First 110-Minute AI Feature Film Made for $2M Open-Sources All Prompts — dr_cintas · 2026-08-11
• MiniMax H3 Video LoRA Training Fails: Unable to Capture Identity — LordDarthShader · 2026-08-11
• Devin AI Builds 3D Models in Blender via Scripting and Screenshot Iteration — davidfromkansas · 2026-08-11
• MiniMax video generation pain point: resolution shift breaks first/last frame transitions — sunshine-3D-Art · 2026-08-11

Multimodal channel →