Pruna AI shrinks MiniMax-H3 from 143 GB to 20 GB and turns the video model into an image generator

Large video models usually live on clusters: running them at home or in a small studio is almost unrealistic. Pruna AI has shown the opposite route. The company took the MiniMax-H3 video model (its original weights take up 143 GB), converted it into an image generator called P-H3-Image, and released it with open weights. The model now weighs about 20 GB, generates an image in 8 steps instead of 50, and needs 21 GB of video memory (VRAM) for an 8-megapixel (MP) frame. Put simply, what used to require a rack of GPUs can now run on a single RTX 4090-class card. The model can generate images from text and edit existing ones.

Why make images from a video model

To produce convincing video, a model has to understand how an object looks from different sides and how light falls on glass, fabric or skin. That knowledge also applies to a single frame, which makes a video model a good foundation for images. But H3 itself produces a "soft" single frame, and it demands a lot of hardware.

Pruna AI reworked the model in several steps:

• The text encoder that reads the prompt was cut from 32B to 4B parameters. Its weight dropped from 67 to 9 GB.
• The number of generation steps was reduced from 50 to 8 using DMD distillation, a method that trains a "fast" model to reproduce the output of a "slow" one. As a result, neither CFG (stronger prompt adherence) nor a negative prompt is needed.
• The timestep block, which accounted for 40% of the transformer, was replaced with a compact equivalent.
• The decoder was retrained: it used to work with video, and now it works with a single frame.

According to the authors, on most prompts the 8-step result is indistinguishable from the 50-step one.

What you need to run it

The minimum is 21 GB of VRAM in w4a8 mode for an 8 MP frame. The RTX 4090 and 5090 qualify, and 24 GB is just enough. With int8 you need 31 GB, and with bf16 you need 56 GB. The int8 kernels require an NVIDIA GPU from the Ampere generation or newer, or an AMD RDNA3/4 card.

ComfyUI support is native, with no custom nodes. For now, though, it is only available in the latest GitHub version (nightly); there is no stable release.

The model accepts up to eight input images. The first is edited, and the rest serve as references: for example, you can show a character or object from different angles. It supports resolutions from 2 to 16 MP, and works best at 8 MP. Below 8 MP there are more artifacts, and 4 MP is a compromise on speed.

According to the authors, the model is strong at documentary and editorial photography, portraits with lifelike skin texture, "cinematic" shots, anime, risograph, and graphics with short text.

Limitations to know about

Consider these weak spots before adopting the model:

• Hands, small figures in the background and fine details "swim," especially below 8 MP.
• Dense text is unreliable. Short quotes in quotation marks come out well.
• Familiar add-ons like "8k, masterpiece" produce an overly glossy image and blown-out highlights. For that reason, a prompt upsampler based on Qwen3.5-9B, which rewrites the request, is enabled by default.
• 4-bit quantization can spoil the composition. If frame accuracy matters, it makes sense to compare the modes.

What to do in practice

If you have a single top-end consumer card, you can test the model on editorial illustrations, concepts and reference-based editing without renting a cluster. For production, it is better to wait for a stable ComfyUI release or pin a specific nightly version. As for the recipe itself — compress the text encoder, cut the steps and retrain the decoder — it is a useful example of how to turn a heavy model into a compact, working tool.

You can try it right away: the project has a demo on Hugging Face, a Pruna AI blog, the weights, and a separate package for ComfyUI.