MiniMax H3 Turbo LoRA Faster Sampling Steps & Prompt Agent Skill — Free Workflows (Patreon)
Downloads
Content
Video : https://youtu.be/4mFGfCWNOdw
Hey everyone!
New video is up: MiniMax H3 Turbo LoRA Faster Sampling Steps & Prompt Agent Skill. Also covers Spectrum, Sol-Attn, and the MiniMax H3 prompt agent skill.
H3 is still climbing Hugging Face. The Turbo LoRA cuts sampling from the old 20-step habit down to 4, 8, or 12 steps with a native Load LoRA node. In the video I also test Spectrum and Sol-Attn as model patches. They help speed, but at 480p they can pixelate on far shots and small details. My general-purpose default stays SageAttention + MiniMax H3 Sigma Shift with Turbo LoRA enabled.
You also see the Hermes agent skill that formats multi-shot prompts from MiniMax's official docs, first-and-last-frame and Ref2V demos, then LTX Upsampler for a cleaner high-res pass.
Free workflows attached:
1. video_minimax_h3_r2v
2. video_minimax_h3_t2vi2v
================================================================================
LORA MODELS REFERENCED
================================================================================
- MiniMax H3 Video Prompt Agent Skill: https://github.com/benjiyaya/Minimax-H3-Prompt-AgentSkill/
MiniMax H3 Turbo LoRA (native ComfyUI-friendly):
- Repo (currently I am using this one in workflow): https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI
- File used in video: 4-step EMA checkpoint 500 pruned comfyui.safetensors
- Workflow files also reference:
minimax-h3\minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors
minimax-h3\minimax_h3_turbo_4step_ema_pruned_comfyui.safetensors
- Purpose: low sampling steps adapter (4 / 8 / 12 steps)
MiniMax H3 Turbo LoRA (custom-node path):
- Repo: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
- Companion nodes: https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
- Purpose: Turbo LoRA with recommended custom-node stack
LTX distilled LoRA (upsampler stage in workflows):
- File: ltx2\ltx-2.3-22b-distilled-lora-384.safetensors
- Purpose: LTX 2.3 second-pass / upsampler path
CharacterSheet Krea 2 LoRA (project reference):
- https://huggingface.co/Alissonerdx/CharacterSheet
- Purpose: character sheet consistency (referenced in project folder)
CUSTOM NODES / RESEARCH LINKS
- Spectrum MiniMax H3: https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
- Spectrum: https://github.com/hanjq17/Spectrum
- Sol-Attn ComfyUI: https://github.com/kijai/ComfyUI-SolAttn_triton
- Sol-Attn: https://nvlabs.github.io/Sana/Sol-Attn/
— Benji