Pruna AI Squeezes a 143 GB Video Model Into 21 GB

Pruna AI has converted MiniMax-H3, a 143 GB video diffusion transformer, into an open-weights image generation and editing model for 24 GB-class GPUs. The resulting P-H3-Image generates images up to 8 megapixels in eight denoising steps, down from 50, and includes native ComfyUI and Diffusers support.

Pruna reports that the smallest package requires a 20 GB download and peaks at 21 GB of VRAM when generating an 8-megapixel image. At 4 megapixels, the model reportedly completes an image in 2.6 seconds, compared with 12.4 seconds for MiniMax-H3. These are vendor benchmarks; performance will vary by GPU, precision, resolution, and inference backend.

A video backbone rebuilt for stills

MiniMax-H3 was built to generate video, so its transformer includes the capacity needed to model changes across frames. Pruna replaced or retrained each major stage, including the prompt upsampler, text encoder, diffusion schedule, transformer weights, and decoder. The new pipeline retains visual patterns learned from video data while specializing the system for text-to-image generation and image editing.

Teams already running MiniMax-H3 have a smaller installation path. A 2 GB LoRA, which contains a compact set of weight adjustments, adapts MiniMax-H3’s FL2VA transformer without requiring another full checkpoint. Pruna specifies a LoRA strength of 1.0 and provides corresponding ComfyUI workflows.

Three memory targets

The release offers three precision tiers that trade memory use against output fidelity. BF16 retains 16-bit floating-point values, INT8 uses 8-bit integers, and W4A8 combines 4-bit weights with 8-bit activations. Components can be mixed, although the published figures cover matched bundles.