Minimax H3 Roundup: Open Source Training Studio, LoRA Patches, and Long Video Workflows

optimisticalish · reddit · 2026-09-01

This Reddit post summarizes Minimax H3 ecosystem updates as of August 31, 2026:

• Inline Studio Update: The open-source local tool now supports training H3 character LoRAs (.char files) with just two images, requiring 24GB VRAM; training on Flux.2 Klein 4B Base needs only 10GB VRAM, and the resulting .char files are portable across models.
• AI Toolkit Patch: A patch for the LoRA trainer adds video alignment, making a control video a true v2v guide instead of a loose reference.
• New LoRAs: Released a 1980s VHS style LoRA and a facial realism close-up LoRA, both with samples.
• Long Video Workflow: Released a ComfyUI workflow bundle for H3 seamless long video chaining, including legacy nodes to prevent breakage from auto-updates.
• Real-time Game Demo: Showcases an 'Interdimensional Game' made with the real-time version of Max, generating every frame on the fly, described as a 'playable film'.

Original post →

More from Multimodal

• Breeze-TTS-2 demo now available on Hugging Face — BreezeBlue · 2026-09-01
• 求助：Anima 模型训练正常但推理生成模糊 — a_throwawayorsmthn · 2026-09-01
• 求测：Ideogram 4 INT8 量化版在 3060 12G 上的表现 — WhyDoiHearBosssMusic · 2026-09-01
• Help: Swapping a Mustache Using a Reference Image in Flux Workflows — sadboi2021 · 2026-09-01
• Fal's new model heralds next chapter for Hollywood VFX — briannekimmel · 2026-09-01
• Sea Angels Generated in p5.js Using Claude Opus 5 — anselm · 2026-09-01

Multimodal channel →