Asirus/MiniMax-H3-Ref2VA-NEXORA

NEXORA H3

NEXORA H3 is a community-created merged MiniMax H3 Ref2VA checkpoint for reference-guided video generation, including audio-video and motion-oriented use cases. I researched and tested different approaches to improve reference consistency, image quality, motion, natural-looking skin, and a restrained cinematic look, then brought those experiments together in one checkpoint. The model is designed to balance these qualities; results still depend on the prompt, references, and workflow.

Access to repository files is controlled by the Hugging Face manual access gate; gated files require owner approval.

Preview videos

These examples were generated with NEXORA H3. They illustrate several motion and audio-video use cases. They are examples, not controlled benchmark results. The clips are also stored in the repository under the examples folder.

1. Cyclist — moving camera and outdoor motion

Open or download the cyclist example

2. Fashion walk — subject and outfit consistency across shots

Open or download the fashion example

3. Singing portrait — audio, lip movement, and gestures

Open or download the singing example

4. Fitness — full-body movement

Open or download the fitness example

Recommended prompt-writing guide

For writing scene prompts for this checkpoint, use NEXORA H3 Ref2VA Motion-Enhanced Prompt Guide. It is the closest match among the six reviewed documents: it is written for MiniMax H3 Ref2VA, uses the six-section Full-Reference CIR format, and gives detailed guidance for reference identity, coherent action, camera choreography, continuity, and audio. The guide is an LLM prompt-writing aid, not a guarantee of model behavior or a replacement for the ComfyUI workflow settings.

Model development

The starting point was the clean Comfy-Org MiniMax H3 Ref2VA INT8 ConvRot diffusion checkpoint. I built NEXORA H3 by merging complementary weight-level changes into this base checkpoint. It is a merged model checkpoint, not a full-parameter fine-tune or a model trained from scratch.

The merge preserved the base architecture and tensor layout. The structural audit found no unmatched targets or shape errors. The INT8 requantization audit measured a median relative error of about 0.549%, a 95th percentile of about 0.679%, and a maximum of about 0.809%. These are tensor-level numerical measurements, not a perceptual benchmark or a guarantee of visual quality.

NEXORA H3 brings these research goals together in one ready-to-use checkpoint. This model card describes the goals and intended use at a high level. The original terms for the base model and any third-party material continue to apply; this project does not grant additional rights.

Recommended ComfyUI setup

• Load the checkpoint as a diffusion model from the ComfyUI models/diffusion_models folder.
• Keep the compatible MiniMax H3 text encoder, VAE, and reference/audio workflow components installed separately; they are not bundled into this diffusion checkpoint.
• Recorded sampler: Euler SDE.
• Recorded scheduler: Simple.
• Recorded sigma shift: 12.3.
• For the best result observed by the creator, use a two-pass workflow. Exact pass settings depend on the workflow and target resolution.
• These settings describe the creator's setup and are not a universal guarantee. A different ComfyUI workflow or node pack may label or expose the corresponding controls differently.

Files

• MiniMax-H3-Ref2VA-NEXORA.safetensors — merged diffusion checkpoint, approximately 19.53 GiB.
• The examples folder contains four audio/video sample clips.
• config.json contains human-readable project metadata and recorded settings. It is descriptive metadata, not a Transformers architecture configuration.
• NOTICE and LICENSE-MINIMAX-H3.txt contain the upstream license notice and MiniMax H3 license text.

Sources and licensing

The base model is Comfy-Org/MiniMax-H3. This derivative remains subject to the MiniMax H3 Community License Agreement. This repository includes a notice identifying the derivative and the upstream license. Any third-party contributions remain subject to their original terms; review those terms before reuse or redistribution. This project does not grant additional rights to upstream materials.

Hub metrics

Downloads, likes, and other repository statistics are reported dynamically by Hugging Face. They are intentionally not hardcoded in config.json, because a stored number would become stale.

Русский

NEXORA H3 — объединённая модель MiniMax H3 Ref2VA для генерации видео по референсному изображению, включая аудио-видео сцены и сцены с движением. Я исследовал и проверял разные подходы к улучшению соответствия референсам, качества изображения, динамики, естественности кожи и сдержанного кинематографичного вида, а затем собрал эти эксперименты в одном checkpoint. Результат зависит от промпта, референсов и workflow.

Модель создана объединением изменений весов с чистой базовой MiniMax H3 Ref2VA INT8 ConvRot. Это merged checkpoint, а не полноценное дообучение на новом наборе данных и не модель, обученная с нуля. Архитектура и структура тензоров базы сохранены. Структурный аудит не выявил несопоставленных целей или ошибок форм. Ошибка повторного квантования INT8: медиана около 0,549%, 95-й процентиль около 0,679%, максимум около 0,809%. Это численная ошибка тензоров, а не оценка визуального качества.

Эта карточка описывает цели и применение модели без технического перечня компонентов сборки. Условия лицензий базовой модели и сторонних материалов продолжают действовать. Для написания сценовых промптов рекомендуется инструкция NEXORA H3 Ref2VA Motion-Enhanced: она рассчитана на MiniMax H3 Ref2VA и подробно описывает работу с референсами, последовательностью действий, камерой, непрерывностью и аудио. Это руководство для LLM, которая составляет промпт, а не гарантия результата.

Зафиксированные настройки: Euler SDE / Simple / sigma shift 12.3. Автор рекомендует двухпроходный workflow для наилучшего результата. Текстовый энкодер, VAE и остальные нужные компоненты MiniMax H3 устанавливаются отдельно. Доступ к файлам регулируется ручной системой допуска Hugging Face; для получения доступа требуется одобрение владельца.