Pruna turns MiniMax H3 video model into an 8-step image generator

Anyone assembling AI video knows the pain: first you need a frame, then you bring it to life. If the frame comes from one model and the animation from another, the style drifts. Lighting, texture and color slip by the second shot, and then manual fine-tuning begins. Pruna offers a workaround: take the MiniMax H3 video model and turn it into a generator of still images. It is called P-H3-Image.

What we know about the release

The information so far comes from a single short post by the enthusiast @alemotion. The model generates an image in 8 steps and works without CFG (classifier-free guidance). This matters for a practical reason: CFG usually requires an additional model pass at every step, so dropping it saves both time and memory.

Editing supports up to 8 input images. You can feed in a character, a location, an object and a style reference at the same time and ask the model to assemble a new frame from them. The model works natively in ComfyUI, so there is no need to write your own wrapper.

The memory figure deserves separate mention. The 4-bit weight set at a resolution of 8 megapixels peaks at about 21 GB of VRAM. The RTX 4090 has 24. The headroom is small, but it exists: a consumer card can handle large frames without renting cloud compute.

Why it matters in practice

The author's scenario: create keyframes that already carry the H3 "look" and only then animate them. The logic is simple. If the starting frame comes from the same model family as the video, the style at input and output should diverge less. This is the author's hypothesis, and the post has no measurements, but the idea is clear and testable.

Imagine an ad spot made of six shots. You upload three references of the hero, one of the environment and one style reference, get a storyboard at 8 steps per frame, and approve it. Only after that do you spend expensive video generations. A mistake at the frame stage costs many times less than a mistake at the video stage.

There is a second advantage: 8 steps without CFG mean fast iterations. For storyboarding that is critical, because you almost always make dozens of variants before finding the right one.

Limitations

There is only one source: a short post with 8 views. There are no comparisons with other image models. It is not shown how many seconds generation takes on a 4090, how quality holds up with 8 references, or how much 4-bit quantization degrades detail compared with full precision.

The 21 GB figure describes the peak at 8 MP. In a real ComfyUI graph, where an upscaler or ControlNet runs alongside, 24 GB will run out quickly. Finally, check the Pruna and MiniMax licenses before any commercial use.

What to do

If you already have an H3 pipeline, take one brief and run it through P-H3-Image and through your current image model. Compare how well the frames match the style of the subsequent animation. If you don't use H3, it is too early to switch for this release: there are no independent tests yet.

But the trend is visible. Video models are starting to get image "twins" so that the whole pipeline from storyboard to animation holds a single visual style. For studios and solo creators it is a chance to cut manual fine-tuning, and for tool developers a signal that consistency between stages is becoming a task in its own right.