FastVideo FastH3 V2 Open Source Release: An Efficient Video Model with 8-Step Generation

A new entry has arrived in the open-source video generation space. The FastVideo team recently published open weights for FastH3 V2 on Hugging Face, centered around an "8-step generation" approach for efficient inference. They also released ComfyUI workflow templates alongside it, further lowering the barrier to local deployment and creative experimentation.

What Is FastH3 V2

FastH3 V2 is the latest iteration of FastVideo's video diffusion model. Its core selling point is baked right into the name: "8-Step" — meaning video generation completes in just 8 sampling steps. Compared to traditional diffusion models that often require dozens of steps, this few-step approach dramatically reduces inference time, making it especially appealing for users with consumer-grade GPUs and creators who value fast iteration.

Model weights are available for download on Hugging Face at FastVideo/FastVideo-FastH3-8-Step-V2. As an open-weight model, users are free to download, deploy, and run it locally — a stark contrast to closed-source solutions that only offer API access.

Out-of-the-Box ComfyUI Integration

One of the highlights of this release is its deep integration with the ComfyUI ecosystem. FastVideo officially provides two complete workflow templates, covering the two most common video generation scenarios:

Text-to-Video (t2v) and Image-to-Video (i2v)

• Text-to-Video: Generate video clips directly from text prompts. The workflow template has been included in the official Comfy-Org repository.
• Image-to-Video: Use a static image as a starting point and let the model animate it — ideal for extending existing visuals into dynamic sequences.

Model weights specifically packaged for Comfy are available in a separate repository: FastVideo/FastVideo-FastH3-Comfy. Users simply download the weights and import the workflow JSON to get started. This "weights + ready-made workflow" combination significantly reduces setup friction, sparing newcomers from the usual headaches of building node graphs and tuning parameters from scratch.

ComfyUI is one of the most widely used node-based workflow tools in the open-source AI image and video generation space. It breaks down model inference into a visual directed acyclic graph (DAG), where each node handles a discrete operation — loading a model, encoding a prompt, sampling, decoding, and so on. Users build complete generation pipelines by connecting nodes, with no coding required. This design makes it easy for non-programmers to customize inference flows, while enabling the community to share and reuse workflows as JSON files. Comfy-Org is the official organization behind ComfyUI; workflows included in their official repository have undergone a degree of quality review, offering users a reasonable baseline of stability and compatibility.

Why Few-Step Generation Matters

FastH3 V2 continues FastVideo's consistent focus on efficiency. Compared to image generation, video generation demands exponentially more compute — the more sampling steps and the longer the clip, the more memory and time it consumes. An 8-step generation approach, achieved through distillation or accelerated sampling techniques, brings inference time down to a manageable range, opening the door for consumer hardware to participate in video creation.

For the open-source community, the value of these efficient models goes beyond raw speed. Lower iteration costs mean the community can test prompts more frequently, tune parameters more aggressively, and explore creative boundaries more freely.

The core techniques behind few-step generation are typically Consistency Distillation or Flow Matching Acceleration. Standard diffusion models perform a full forward pass through the neural network at each denoising step — the more steps, the heavier the computation. Distillation trains a "student model" to mimic the final output of a "teacher model" that takes many steps, allowing the student to jump close to the target in very few steps. Flow matching reparameterizes the diffusion process mathematically, making the trajectory from noise to image/video more linear, which naturally supports fewer integration steps. Both approaches inevitably involve a quality-speed tradeoff: too few steps, and the model struggles to correct early errors, potentially resulting in blurry details or unnatural motion. The choice of 8 steps typically reflects an empirically validated balance between efficiency and quality — not an arbitrary number.

How to Get Started

For users eager to try it out, the basic workflow looks like this:

1. Download the model weights from FastVideo-FastH3-Comfy on Hugging Face;
2. Grab the official t2v or i2v workflow JSON file;
3. Import the workflow into ComfyUI and load the corresponding weights;
4. Enter a text prompt or starting image and run the generation.

The FastVideo team (haoailab) has also shared generation examples on social media, so users can get a visual sense of the model's real-world output before diving in.

Wrap-Up

The open-source release of FastH3 V2 is a snapshot of the ongoing evolution in open-weight video models. With 8-step generation at its core and solid ComfyUI workflow support, it takes local video generation one step further in practical usability. For creators and developers following open-source AI video tools, this is a new option worth adding to the testing list.

It's worth noting that this article is based on official release information. Actual output quality, stability, and hardware requirements remain to be validated through broader community testing.