Related article

MiniMax H3 | Singularity V2: Enhanced Motion, Prompt Adherence, and Style

A comprehensively fine-tuned V2 INT8 release that the author says expands motion coverage and improves prompt adherence, audio references, and style performance.

English edition published · Latest update · Sources checked 5 Nov 2026 · Chinese original ↗

Verified on: 2026-11-05. The repository update time is not the same as the model release date.

Video generation > Fine-tuning · Motion control / Style / Photorealism / ComfyUI / Audio and video

Guide to the Author’s Model Card

The author positions Singularity as a combined video fine-tuning and model-merging release for MiniMax H3, covering text-to-video, image-to-video, reference-based generation, and video-to-video workflows. V2 is additionally provided in INT8 format. According to the model card, V2 expands its coverage of motion scenarios and movement sequences, improves adherence to descriptions of subjects, actions, environments, and visual styles, and fixes issues with audio reference conditioning. The author lists photorealism, East Asian photorealism, 3D CG, and cinematic CG among the supported styles. These are all claims by the author; Lumobrief has not downloaded the weights or reproduced the examples.

Base Model, Input Modes, and Dependencies

The repository identifies MiniMaxAI/MiniMax-H3 as the base model and primarily targets ComfyUI. The author says the model combines checkpoints including ref, fl, and b25-49 and undergoes high-step fine-tuning, but the model card excerpt does not provide complete training data, recipes, or evaluations sufficient for independent verification. Inputs may include text, a first frame or image, reference materials, or a source video, while the generated output remains video. This information does not establish that the model has standalone image-generation capabilities.

License, Author Claims, and Items Requiring Verification

The page metadata lists Apache-2.0, but the base model is also subject to the MiniMax H3 Community License Agreement. Users should therefore review both the repository files and the upstream terms. The claimed audio reference “fix” and improvements to motion and style still require independent testing. The author’s recommendation of a prompt length of approximately 5,000 characters is not a universal guarantee. Source: FenomAI/Minimax-h3_Singularity. Verified against revision bc1d190b; the repository update on 2026-11-04 represents only the repository modification time and is not the V2 release date.

Article Update Log

The initial listing and subsequent documentation revisions are recorded below. No independent performance testing was conducted.

Return to the MiniMax H3 version directory

Creator demos

These are demonstrations supplied by the creator, not independent tests by Lumobrief.

Author model card · 1.mp4 · Original source ↗
Author model card · 3.mp4 · Original source ↗

Sources and repositories