Video Rebirth compresses 31B video-audio model to 14B parameters

Video Rebirth has released HyperFlow H3 Flash, a highly compressed version of the MiniMax-H3 model that enables joint video and audio generation. This development matters because it significantly reduces the computational resources required to produce synchronized audiovisual content, making high-quality generation more accessible on consumer-grade hardware.

What is HyperFlow H3 Flash?

According to the developers at Video Rebirth, HyperFlow H3 Flash is a compressed, few-step version of the MiniMax-H3 transformer. While the original MiniMax-H3 model contains approximately 31 billion parameters, the HyperFlow version has been reduced to roughly 13.7 billion to 13.9 billion parameters, depending on the specific task being performed.

The model is designed for the simultaneous generation of video and synchronized stereo audio. It can produce video clips ranging from 5.2 to 15 seconds at 24 frames per second, with a default resolution of 1344×768. The accompanying audio is delivered as a 32 kHz stereo soundtrack.

How does the compression affect performance?

The primary advantage of this compression is the speed and efficiency of the sampling process. The model can generate output in as few as 4, 6, or 8 steps without requiring classifier-free guidance (CFG).

Performance benchmarks provided by the source indicate varying speeds based on the hardware used. For a 5.2-second clip using 6 steps, the model takes approximately 8.7 seconds on four B200 GPUs, 36 seconds on a single H200, and 92 seconds on a single NVIDIA RTX 5090. Notably, the developers stated that the model can run in real time on four Blackwell GPUs when used with NVIDIA SOL-H3, an inference-acceleration partner.

What are the hardware requirements for users?

Because of the compression and the availability of different formats, the hardware requirements are flexible. The model is available in GGUF and Python INT8 formats.

For users on personal computers, the developers suggest using the GGUF format if the system has 64 GB of RAM or less. For those using data-center GPUs or high-end consumer cards like the RTX 24 GB or 32 GB series, the INT8 version is recommended. The smallest footprint for the model is approximately 9.1 GB in Q4_K_M precision, while the Q8_0 version requires about 14.9 GB. This allows the model to run on consumer GPUs with 24 GB of VRAM while utilizing 16 GB of system RAM.

Which tasks can the model perform?

The HyperFlow H3 Flash architecture supports three primary input-to-output tasks:

1. Text-to-Video and Audio (t2va): Users provide a text prompt to generate a video with sound.
2. First Keyframe-to-Video and Audio (fl2va): This allows for input via a prompt combined with a first keyframe, or both a first and a last keyframe, to guide the generation.
3. Reference-to-Video and Audio (ref2va): This more complex task allows users to input a prompt along with up to 12 references, which can include images, videos, and audio clips.

How can developers implement this model?

Developers can integrate the model through several runtimes. It is compatible with ComfyUI via custom nodes, Unsloth Studio, and stable-diffusion.cpp. For those preferring a programmatic approach, a Python package is available that supports the diffusers library.

When using the Python package, the first run requires a significant download of approximately 90 GB, though this can be reduced to roughly 68 GB if the text encoder precision is set to fp8. The model supports various offloading modes—such as 'auto', 'model', and 'stream'—to manage GPU memory placement effectively depending on the available VRAM.

How does this relate to creative workflows?

While HyperFlow H3 Flash is an open-weights model intended for local or specialized deployment, the underlying technology of efficient, multimodal generation is central to the future of digital creation.

For creators looking to experiment with these types of generative capabilities immediately, Grafy Creator provides a streamlined workspace for generating high-quality images and video from prompts or photos. While users may not be running raw 13B parameter diffusion models on their local machines through Grafy, the goal of rapid, high-fidelity visual output remains a core part of the creative process.

Frequently asked questions

Does HyperFlow H3 Flash generate audio?

Yes, the model is designed for joint video and audio generation. It produces synchronized stereo audio at a 32 kHz sampling rate alongside the generated video clips, ensuring that the sound and visuals are matched.

How many steps are needed for generation?

Users can choose between 4, 6, or 8 sampling steps. The 4-step setting is optimized for speed, the 6-step setting is the default for balanced performance, and the 8-step setting is intended for higher quality outputs.

Can this model run on a consumer GPU?

Yes, the model is optimized to run on consumer hardware. Specifically, it can run on NVIDIA RTX GPUs with 24 GB or 32 GB of VRAM, provided the system has at least 16 GB of system RAM to support the GGUF or INT8 formats.

---

Source: Hacker News — front page — Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8