Pika H3 Flex: Open video model with sound in just six sampling steps

What it's about

Pika Labs has released Pika H3 Flex on Hugging Face. It is a post-trained model built on MiniMax-H3 that generates video with synchronized stereo sound. The selling point, according to the vendor: it needs only six sampling steps (six passes through the model per generation), where other video models often need far more. Pika calls it a tool for professional creators who want to iterate quickly with precise control. The weights, code and ComfyUI workflows are all public. The release surfaced in a Reddit post on r/StableDiffusion.

Structure: two adapters on one base model

Pika H3 Flex is a video-and-audio model family with two "task adapters." An adapter here is a LoRA: a small add-on layer that steers a large base model toward a task without replacing all of its weights. Both adapters ship as FP32 LoRAs with rank/alpha 128.

• FL2VA (text and endpoint images): text-to-video, first frame, last frame, or both.
• Ref2VA (multimodal references): the model receives ordered image and video references, plus optional audio.

The pipeline encodes the prompt and input media, generates video and audio latents (compressed intermediate representations) with the chosen adapter, decodes them and muxes everything into a single MP4 file. The repository also contains the MiniMax-H3 base components in Diffusers format (transformer, text encoder, VAE, audio VAE, tokenizer). The transformer is only the base model; you have to apply the matching Pika adapter separately. One note from the manifest: generic six-step Euler sampling is not equivalent. You must use the bundled, exact sampling schedule.

Technical specs

Duration is rounded up to the model's frame grid: 5 seconds gives 124 frames (about 5.17 s), and 15 seconds gives 362 frames (about 15.08 s).

References and modes

Supported modes are text-to-video, first image (i2va), last image (l2va), both images (fl2va) and reference mode (ref2va). Ref2VA takes up to 9 images, 3 videos and 3 separate audio clips, 12 files in total; at least one image or video is required. Videos and audio clips must each be 2–15 seconds long, and the total of the videos and the total of the audio clips may not exceed 15 seconds each. A video's embedded audio does not count as a separate audio file. Pika stresses that these are only accepted input limits. Not every maximum combination fits into GPU memory.

The example prompt in the repo shows what this means: a table tennis scene in an empty pool with camera moves timed to the second (top-down, crane shot, pans, a 180-degree orbit), characters kept consistent via a reference image, and detailed descriptions of sound effects and music. Pika notes that the showcase clips are creative examples, not validation renders or throughput benchmarks. Exact reproduction isn't guaranteed, and seeds and preprocessing are unspecified.

Pika API extensions

Separately from the downloadable weights, Pika offers hosted services: 1080p, 2K and 4K video from a draft video, separate audio stems, and prompt enhancement. These features don't apply to local inference, which delivers native 480p/768p with mixed stereo sound.

What you need to try it

The base components take up about 210 GB; a workflow loads only the transformer partition that matches it. The Pika CLI runs one request on one GPU. Representative 480p and 768p requests were tested on an 80 GB H100. Large requests need --cpu-offload and plenty of host RAM. That puts it at workstation or cloud level rather than gaming-PC level. The source gives no concrete figures for smaller cards (details are in the repo's hardware docs). You can install via pip, Conda or Docker; the pika-h3-pro command downloads the base and adapters automatically. Alternatively, there are ComfyUI workflows (T2V, FL2VA, R2V) with BF16 LoRAs, plus Diffusers examples. Google Colab and Kaggle are mentioned as platforms, but whether their hardware is enough remains unclear.

License

The adapters are derivatives of MiniMax H3 and fall under the MiniMax H3 Community License. The bundled base components keep the MiniMax license, the Qwen encoder and tokenizer are Apache 2.0, and so is the example code (but not the weights or sample media). Pika asks for attribution and a citation of the "Technical Deep Dive."

Our take

Six steps for video with sound is an attractive speed promise. Reference control with up to 12 files and prompt-based camera direction are also handy for multi-scene storytelling. So far, though, the only evidence is Pika's own showcases. Independent tests, speed measurements and quality comparisons are missing from the sources. Running it locally remains a high bar (210 GB base, an 80 GB card as the test configuration), and higher resolutions are tied to the paid Pika API. The community license may include usage restrictions, so check it before any commercial use.