PickP-H3-Image: 8-Step, CFG-Free 8MP Image Generation and Editing

A text-to-image generation and editing model based on MiniMax-H3. It renders images at 8-megapixel resolution in just 8 steps without classifier-free guidance (CFG). It excels in photorealism, film and editorial styles, and graphic design featuring short quoted text.

1 / 7

A Qwen3.5-9B-based prompt upsampler is enabled by default, converting one-line requests into the optimal form for the model. For editing tasks, it accepts up to 8 input images, modifying the first image while using the rest as reference images. For example, requests such as 'Change the woman's hat in photo 1 to the hat in photo 2' are possible.

It includes a 33.1B parameter transformer and a 4B text encoder, with a total size of 100GB in bf16. When using the prompt upsampler, approximately 101GB of GPU memory is required at 8MP resolution, but it can run on an 80GB GPU with offloading. Text-to-image generation takes approximately 9 seconds on a B200.

It is sensitive to prompts, so avoid exaggerated adjectives; the first sentence clearly specifying the image type is important. Artifacts in faces or hands may increase at resolutions below 8MP, and video or audio generation is not supported. int8 and 4-bit files for ComfyUI are provided in a separate repository.