MiniMax character LoRA on Runpod: 150 images, ~2300 steps to nail likeness, and 5090 beats H100

Draufgaenger · reddit · 2026-09-04

A practitioner shares hands-on findings from training MiniMax character LoRAs on Runpod (with a 2-minute video tutorial):

• Datasets: image-only datasets worked much better than mixed ones; focus on near-frame-filling, sharp faces; MiniMax needs more images than LTX/Wan (150 images for a blonde woman's likeness, 50 sufficed for the author's own face);
• GPU surprise: RTX 5090 beat the H100 on this image-only workload; RTX 6000 PRO was only 20% faster than the 5090;
• Step counts: the author's face peaked at 700 steps; the woman's dataset peaked at 2310 steps — and notably dataset size didn't move the peak (2300 steps regardless of 50 or 150 images), though the 150-image LoRA was objectively much better;
• Data noise matters: 4 brown-haired images out of 150 caused brown-haired outputs past the peak even with blonde prompts;
• Natural-prose captions including lighting and look work best; voice training isn't supported in diffusion pipe, but MiniMax accepts a reference voice.

Original post →

More from Multimodal

• Open-sourced pixel art normals generator ships as Rust CLI and wasm library — HugoDuprez · 2026-10-20
• Fibo 2.0 is strongest for ecommerce and creator content, weakest at UIUX design — ArtificialAnlys · 2026-10-20
• Fibo 2.0 leads old Fibo on all 9 capabilities, best at reasoning and anatomy — ArtificialAnlys · 2026-10-20
• Fibo 2.0 ranks #6 on enhancement & restoration, beats Fibo Edit on all 7 actions — ArtificialAnlys · 2026-10-20
• Bria's Fibo 2.0 tops open-weights image editing leaderboard at $32 per 1k images — ArtificialAnlys · 2026-10-20
• 'The Queen in Yellow' Part Two: AI Horror-Comedy Animation Goes Viral on Reddit — Cappinski · 2026-10-20

Multimodal channel →