PrunaAI releases P-H3-Image: 8 MP image generation and editing in 8 steps, built on MiniMax-H3

PrunaAI has released P-H3-Image, an image model for text-to-image generation and image editing that renders 8-megapixel images in 8 steps without classifier-free guidance (CFG). The model is based on MiniMax-H3 and is published on Hugging Face as a diffusers pipeline.

Specifications

• Tasks: text-to-image, and image editing with 1 to 8 input images
• Sampling: 8 steps, no CFG, no negative prompt
• Resolution: 8 MP for text-to-image, 4 or 8 MP for edits
• Weights: 33.1B-parameter transformer, 4B text encoder, VAE and prompt upsampler, 100 GB in bf16
• GPU memory at 8 MP: about 101 GB with the prompt upsampler, 83 GB without
• Prompt upsampler: included and on by default, based on Qwen3.5-9B
• Pricing: not yet disclosed

Hardware requirements

At 8 MP in bf16, the model needs a GPU with about 101 GB of memory, such as one H200 or B200. PrunaAI says it can run in 80 GB with offloading. A separate repository, PrunaAI/P-H3-Image-Comfy, provides ComfyUI files in int8 and 4-bit that run in 21 GB of VRAM.

Editing

The same pipeline picks the task from its inputs: passing an image triggers an edit. With multiple images, the first is the one being edited and the others are references that a prompt can name as Picture 2, Picture 3, and so on. Edits work best at 4 or 8 MP. Without explicit height and width, an edit is generated at the size of Picture 1, scaled up to at least 4 MP.

According to the model card, a short request typed as-is often makes the model redraw the whole scene. An edit prompt that names what to keep changes only the requested part. The upsampler writes that form automatically, so a request such as "make the pickup truck red" is enough.

Prompt upsampler

PrunaAI states that P-H3-Image is sensitive to prompting and renders full, specific prompts much better than short requests. The pipeline therefore runs the Qwen3.5-9B-based upsampler on every request. It writes a detailed prompt for the given aspect ratio or, for edits, a brief of what to keep and what to change. The rewritten prompt is returned as a pipeline output, and the step can be removed from the pipeline blocks to render a prompt as written.

For hand-written prompts, the model card recommends opening with one sentence naming the real object the image is (for example, "a two-color risograph print"), describing a single scene with a light source and direction, and quoting every piece of text exactly. It also advises dropping hype words such as stunning, epic, cinematic, vibrant, hyper-detailed, 8k and masterpiece, which it says push output toward a glossy, oversaturated, airbrushed look.

Resolution guidance

PrunaAI says the model works best at 8 MP. Images at 2 and 4 MP can be good, but artifacts become more frequent, especially in faces, hands and small details. The card lists 8 MP text-to-image sizes for seven aspect ratios, with a default of 2912 x 2912 (1:1). Its examples include 3360 x 2496 (4:3), 3552 x 2368 (3:2) and 3872 x 2176 (16:9). Sides are rounded to multiples of 32.

PrunaAI describes the model as strongest at photorealism, film and editorial looks, and graphic styles with short quoted text. No benchmark scores are published on the model page.

What this means

The 8-step, no-CFG sampling design reduces the per-image sampling work at a native resolution of 8 MP. The memory cost is substantial: about 101 GB in bf16 with the upsampler, or 83 GB without it. The 21 GB ComfyUI int8 and 4-bit files make the model reachable on far smaller hardware, though the model page does not give quality comparisons for those formats.

The bundled upsampler is a core part of the pipeline rather than an add-on. PrunaAI's guidance indicates that output quality depends heavily on prompt structure, and the default setup handles that rewriting automatically. Teams that need exact prompt control must remove the step manually.

Up to 8 reference images in a single edit call puts multi-image composition in the same pipeline as text-to-image generation, with the first image as the edit target.