license: other license_name: minimax-h3-community-license-agreement license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE pipeline_tag: text-to-video base_model:

• MiniMaxAI/MiniMax-H3
• Comfy-Org/MiniMax-H3 tags:
• minimax-h3
• text-to-video
• text-to-audio
• audio-video
• synchronized-audio-video
• lora
• comfyui
• four-step
• experimental language:
• zh
• en inference: false

MiniMax H3 Turbo 4-Step — 10Eros Pruned Curve Compatibility LoRA (EXP v001)

An exact-checkpoint-specific, four-step ComfyUI acceleration LoRA for the local 10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors MiniMax H3 fine-tune.

This repository contains an independent LoRA, not a merged 40GB diffusion checkpoint. It keeps the 208 already compatible attention/MLP adapters from a legacy Standard MiniMax H3 Turbo LoRA and converts its 51 full-width AdaLN adapters to the target checkpoint's 8-dimensional pruned curve basis.

[!CAUTION] This is an Experimental, exact-base compatibility build. It is not an official MiniMax release, not a newly trained or re-distilled Turbo model, and not a generally compatible LoRA for arbitrary pruned checkpoints. Use it only with the exact main-model SHA-256 documented below.

[!IMPORTANT] MiniMax H3 and Model Derivatives are governed by the MiniMax H3 Community License Agreement. That agreement defines excluded territories, redistribution conditions, commercial terms, downstream safeguards, an acceptable-use policy, and mandatory LICENSE/NOTICE handling. A public Hugging Face repository can be reachable from excluded territories; repository gating is not necessarily geographic access control. Confirm that you have the right to distribute both this derivative and the 10Eros fine-tuned-base-dependent adaptation. This model card is not legal advice.

中文说明

这是为以下唯一指定主模型制作的 MiniMax H3 四步加速 LoRA：

10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors
SHA-256: f82cc3f723b080e7ae94a7c98f95aa989e387618d0bdc940133dfbd9f432c062

普通完整 MiniMax H3 Turbo LoRA 有 259 组 adapter，其中 208 组 attention/MLP 与该 pruned 主模型形状兼容，另外 51 组 AdaLN 的输入维度为 2688，而主模型只接受 8 维曲线坐标，直接用 Bypass 加载会发生维度错误。

本模型没有简单删除这 51 组 AdaLN，而是：

• 保留 208 组原兼容 adapter，416 个 tensor 与来源逐张量完全一致；
• 将 51 个 AdaLN-A 从 [16,2688] 投影为 [16,8] BF16；
• 51 个 AdaLN-B 保持逐张量完全一致；
• 增加 51 个 FP32 .diff_b，保留仿射投影中的常数响应；
• 最终为 259 组 A/B adapter + 51 个 bias diff，共 569 tensors。

Model details

Weight provenance

The acceleration source is the exact local file:

File: minimax_h3_turbo_4步加速_comfyui.safetensors
SHA-256: 35946f9f2957c2766e28b627c88169535249dd07a3040ce3c2c8c99951fdbc7b

Its embedded conversion metadata points to local source SHA-256 9344cd958f8d354da03dd00b7d462933eb5d0cbf11e56a25d8e9911bb971160e. Independent provenance inspection identifies it as a small BF16/T8 perturbation of the initial/legacy non-EMA Standard four-step line from larryvrh/MiniMax-H3-Turbo-Lora.

It is not v4-step600, ckpt500, ckpt850, or EMA. Do not rename or advertise this file as one of those checkpoints. The current upstream repository may recommend newer weights; that recommendation does not change this file's provenance.

Exact compatible main model

File: 10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors
Size: 40,225,724,112 bytes
SHA-256: f82cc3f723b080e7ae94a7c98f95aa989e387618d0bdc940133dfbd9f432c062
Structure: 532 tensors; adaln_t_table [1025,8]; 50 blocks + 2 token-refiner blocks

The main model is not included here. This card does not establish the authorship, license, or redistribution rights of that local fine-tuned checkpoint. The publisher must verify those rights separately.

What was changed

The full H3 AdaLN path consumes the post-SiLU time embedding

S(t) = SiLU(TimeEmbedder(t)) ∈ R^2688

See the full description on the original page