Home Creators Posts Import Register Favorites Logout

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

Files

MiniMax H3 Turbo LoRA Faster Sampling Steps & Prompt Agent Skill

This video covers MiniMax H3 Turbo LoRA for low-step audio-video sampling in ComfyUI, plus Spectrum and Sol-Attn model patches for faster inference. You also see SageAttention, the native MiniMax H3 Sigma Shift node, first-and-last-frame and reference-to-video demos, and an agent prompt skill built from MiniMax's official prompt guides. The close shows Upsampler pushing H3 output toward 2K and 4K with detailer enhancement. This is for ComfyUI users already running MiniMax H3 who want fewer sampling steps without rebuilding their whole graph. It also fits creators who write long multi-shot prompts, use Hermes or other agent harnesses, and want a repeatable Ref2V / text-image-to-video workflow with Turbo LoRA at 4, 8, or 12 steps. MiniMax H3 hit top trending on Hugging Face fast, and Turbo LoRA cuts the old 20-step habit down to 4–12 steps with native Load LoRA. Spectrum and Sol-Attn help speed, but at 480p they can pixelate on far shots and small details, so the practical default here is SageAttention plus Sigma Shift with Turbo LoRA. Pair that with an agent skill for H3-formatted prompts and LTX upsampling, and you get a local stack that stays usable while you iterate. Freebie Workflows: https://www.patreon.com/aifuturetech/posts/165937570?utm_source=youtube&utm_medium=video&utm_campaign=20260806 MiniMax H3 Video Prompt Agent Skill https://github.com/benjiyaya/Minimax-H3-Prompt-AgentSkill/ Turbo LoRA, thanks Larryvrh and drbaph! https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI Spectrum-style spectral feature forecasting for ComfyUI's native MiniMax H3 audio-video model. https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3 Spectrum https://github.com/hanjq17/Spectrum Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification https://github.com/kijai/ComfyUI-SolAttn_triton Sol-Attn https://nvlabs.github.io/Sana/Sol-Attn/ Timeline : 00:00 - Introduction to MiniMax H3 Turbo LoRA Benji introduces the new MiniMax H3 Turbo LoRA models, which are trending on Hugging Face for their ability to significantly speed up video generation by reducing sampling steps. 00:19 - Exploring Performance Optimization Techniques A look at various methods to enhance inference speed, including ComfyUI custom nodes like Spectrum and Nvidia's Soul Attention, though Benji notes some potential for pixelation with these specific tools. 01:05 - Integrating Turbo LoRA with ComfyUI Benji explains how to use the Turbo LoRA model natively in ComfyUI using the standard "Load LoRA" node, highlighting that it allows for high-quality video in as few as 4 to 8 steps. 02:30 - Native Node Setup and Sigma Shift A walkthrough of the workflow, emphasizing the importance of updating ComfyUI to access the native MiniMax H3 Sigma Shift node for optimal results. 03:33 - Testing First and Last Frame Generation Benji demonstrates how the model handles camera and character motion when provided with specific starting and ending images. 04:32 - The AI Agent Text Prompt Skill Benji introduces a custom "agent skill" he developed. Based on official documentation, this tool helps AI agents generate highly detailed, multi-shot text prompts optimized for MiniMax H3. 06:35 - Practical Demo: Futuristic Fight Scene A step-by-step example using the agent skill to create a 10-second futuristic fight sequence from two reference images. 07:47 - Reference-to-Video Workflow Demonstrating how to guide the AI to identify specific characters and actions (like a racing scene) to create a coherent narrative flow. 08:53 - Final Recommendations and Upscaling Benji shares his preferred setup (Sage Attention + Sigma Shift) for general use and explains why he uses the LTX upsampler to achieve smooth 2K or 4K final outputs. Local Workstation GPU : https://amzn.to/3XfXsAO -------------------------------------------------------------------------------------------------------------------------------- If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech

Comments

I Cant Even Productions

Any idea why I might be getting an error saying my comfyui-LTXVideo pack has an incompatibility issue with my current environment. I got all the files installed but these two ltx nodes are X'ed out "LTXVLoopingSampler" and "LTXVSpatioTemporalTiledVAEDecode" in the subgraph area.

Innovate Futures @ Benji

Comfyui_LTX node from LTX team😳. you haven't install this? I posted LTX workflow before Minimax H3 also using this node, remembered you have asked about LTX workflow there.