Revise NEXORA H3 model card overview

e7b864d verified 10 days ago

9.52 kB

---
license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: Comfy-Org/MiniMax-H3
pipeline_tag: image-text-to-video
library_name: comfyui
tags:
  - minimax-h3
  - ref2va
  - video-generation
  - audio-video
  - int8
  - convrot
  - merged-checkpoint
  - nexora-h3
---
# 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

<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00030_audio.mp4" controls playsinline preload="none" width="360"></video>

[Open or download the cyclist example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00030_audio.mp4)

### 2. Fashion walk — subject and outfit consistency across shots

<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00031_audio.mp4" controls playsinline preload="none" width="360"></video>

[Open or download the fashion example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00031_audio.mp4)

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

<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00032_audio.mp4" controls playsinline preload="none" width="360"></video>

[Open or download the singing example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00032_audio.mp4)

### 4. Fitness — full-body movement

<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00016_audio.mp4" controls playsinline preload="none" width="360"></video>

[Open or download the fitness example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00016_audio.mp4)

## Recommended prompt-writing guide

For writing scene prompts for this checkpoint, use [NEXORA H3 Ref2VA Motion-Enhanced Prompt Guide](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/NEXORA_H3_Ref2VA_Motion_Prompt_Guide_EN.md). 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](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/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](https://huggingface.co/Comfy-Org/MiniMax-H3). This derivative remains subject to the [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). 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](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/NEXORA_H3_Ref2VA_Motion_Prompt_Guide_EN.md): 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.

	---
	license: other
	license_name: minimax-h3-community-license-agreement
	license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
	base_model: Comfy-Org/MiniMax-H3
	pipeline_tag: image-text-to-video
	library_name: comfyui
	tags:
	- minimax-h3
	- ref2va
	- video-generation
	- audio-video
	- int8
	- convrot
	- merged-checkpoint
	- nexora-h3
	---

	# 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

	<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00030_audio.mp4" controls playsinline preload="none" width="360"></video>

	[Open or download the cyclist example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00030_audio.mp4)

	### 2. Fashion walk — subject and outfit consistency across shots

	<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00031_audio.mp4" controls playsinline preload="none" width="360"></video>

	[Open or download the fashion example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00031_audio.mp4)

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

	<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00032_audio.mp4" controls playsinline preload="none" width="360"></video>

	[Open or download the singing example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00032_audio.mp4)

	### 4. Fitness — full-body movement

	<video src="https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00016_audio.mp4" controls playsinline preload="none" width="360"></video>

	[Open or download the fitness example](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/resolve/main/examples/MiniMax_H3_test_00016_audio.mp4)

	## Recommended prompt-writing guide

	For writing scene prompts for this checkpoint, use [NEXORA H3 Ref2VA Motion-Enhanced Prompt Guide](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/NEXORA_H3_Ref2VA_Motion_Prompt_Guide_EN.md). 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](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/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](https://huggingface.co/Comfy-Org/MiniMax-H3). This derivative remains subject to the [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). 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](https://huggingface.co/Asirus/MiniMax-H3-Ref2VA-NEXORA/blob/main/NEXORA_H3_Ref2VA_Motion_Prompt_Guide_EN.md): 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.