MiniMax-H3 Turbo LoRA¶

Development line: project:minimax-h3-turbo-lora · thread minimax-h3-turbo-lora
Last event: 2026-08-06 · 1 dated since 2026-08-06 · Researched: 2026-09-05 · confidence: high

What it is¶

MiniMax-H3 Turbo LoRA is a community adapter with ComfyUI nodes for MiniMax-H3 text-to-video and image-to-video workflows.

• Joint audio-video sampling with fewer steps.
• Base support for full and pruned H3 models.
• Custom sampler for separate audio and video schedules.

Practical sampling takes 4 steps minimum, with 6–8 steps preferred; audio and intense motion remain a preview. Use it for fast iteration, but validate fast-motion shots against the base workflow.

Development line¶

• 2026-08-06 — MiniMax-H3 Turbo LoRA resource bundle documented. On 2026-08-06, documentation connected the MiniMax-H3 Turbo LoRA model resource with a demo Space, a ComfyUI integration repository, and a ComfyUI-oriented model resource. This establishes an accessible workflow across model distribution, demonstration, and ComfyUI use. The linked resources alone do not establish their individual release dates, versions, authorship, or performance claims.

What changed¶

• 2026-08-06 — The public ComfyUI integration added the Turbo LoRA loader and 4-step sampler, then added pruned/curve-base support, runtime application of the adapter, and a low-VRAM switch.
• 2026-08-07 — The v4 step-600 EMA checkpoint became the default recommendation; guidance narrowed practical sampling to 4–8 steps and retained v1 step-850 only for heavy motion at four steps.
• 2026-08-08 — The node integration fixed an AdaLN injection mismatch for audio-reference conditioning.
• 2026-08-10 — The integration corrected modulation-segment calculation through the AdaLN time-table lookup.
• 2026-08-12 — The integration merged an audio-reference AdaLN-row fix.
• 2026-08-14 — The integration merged a further upstream contribution; keep the custom node current.

How to use this¶

From 2026-08-06, evaluate MiniMax-H3 Turbo LoRA through the linked model and ComfyUI integration path rather than as an isolated checkpoint; verify live compatibility, licensing, and setup requirements before use.

1. Install Larryvrh/ComfyUI-MiniMax-H3-Turbo with ComfyUI-Manager or clone it into ComfyUI/custom_nodes, then restart ComfyUI. — https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
2. Download the Turbo LoRA .safetensors into ComfyUI/models/loras/ and install the MiniMax-H3 base model, VAEs, and text encoder. — https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
3. Start from an official H3 text-to-video or image-to-video workflow; place the Turbo LoRA node between the diffusion-model loader and SamplerCustomAdvanced, then feed that sampler from MiniMax-H3 Turbo Sampler. — https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
4. Set the scheduler to simple, strength to 1.0, and begin at 6–8 steps; use 4 only when speed matters more than motion quality. — https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

Best practices¶

• Use minimax_h3_turbo_v4_step600_ema.safetensors for most shots; reserve v1 step-850 for unusually fast, heavy motion at exactly four steps. — https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
• Keep low_vram off for the sharper runtime application; enable it only after an out-of-memory failure, accepting softer output on quantized bases. — https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
• Keep the Turbo Sampler in the graph across ComfyUI updates because it adapts to native ModelSamplingAV support and preserves correct audio scheduling on older builds. — https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo

Superseded by this¶

• 2026-08-07 — v1 step-850 is no longer the general default; v4 step-600 EMA is preferred except for four-step heavy-motion clips.
• 2026-08-06 — Earlier guidance limited the adapter to non-pruned H3 bases; the custom node added support for pruned and curve variants.

Still unknown¶

• The adapter and its node are community releases, not an official MiniMax Turbo product.
• No independent benchmark was found that establishes a universal quality or speed ranking across GPU, resolution, duration, and motion type.

Sources¶

Agent brief¶

• Subject: project:minimax-h3-turbo-lora, thread minimax-h3-turbo-lora, 1 dated events 2026-08-06 → 2026-08-06.
• Practical note: From 2026-08-06, evaluate MiniMax-H3 Turbo LoRA through the linked model and ComfyUI integration path rather than as an isolated checkpoint; verify live compatibility, licensing, and setup requirements before use.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.