P-Image-OSS
Open-source image generation and editing model based on MiniMax-H3
2 min readupdated
P-Image-OSS is an advanced model for text-to-image generation and comprehensive image editing, created by PrunaAI based on the MiniMax-H3 architecture. The model produces detailed images at up to 8 megapixels in just 8 generation steps without a classifier-free guidance (CFG) scale. A key feature of the pipeline is a built-in text upsampler (prompt enhancement module) based on Qwen3.5-9B, which automatically turns a short request into an expanded, detailed prompt. The new model excels in photorealism, cinematic, and editing styles, as well as in creating graphics with short quoted text.
Key features
- Fast generation: A full image creation cycle takes only 8 sampling steps without the need to configure CFG and negative prompts.
- High resolution: Native support for generating images at 8 MP quality (for example, 2912×2912 pixels) with fine detail in skin and textures.
- Multi-component editing: The ability to edit a single image or combine up to 8 different references into a single scene using a text prompt.
- Built-in upsampler: Automatic enhancement of user queries using a language model to achieve the best visual result.
What you can do
- Text-to-image: Generate a photorealistic frame, illustration, or high-resolution poster with precise proportions and short text adherence.
- Targeted editing: Modify specific elements in an existing photo by adding or removing details using a text instruction.
- Reference-based compositing: Combine characters and objects from multiple different images (up to 8) into a single cohesive composition.
- Pipeline integration: Run the model locally using the Diffusers library to automate content generation.
Example
Example of generating an image from a text prompt using Python and Diffusers:
import torch
from diffusers import ModularPipeline
pipe = ModularPipeline.from_pretrained("PrunaAI/P-Image-OSS")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
prompt = "A cinematic street view at night with glowing lights."
image = pipe(
prompt=prompt,
height=3360, width=2496,
generator=torch.Generator().manual_seed(3),
output="images"
)[0]
image.save("output.png")Read more
Details• Tasks: Text-to-image and editing using 1–8 images
• Architecture: 33.1 billion parameter transformer, 4 billion parameter text encoder
• VRAM requirements: About 101 GB with the prompt upsampler (in bf16), or from 80 GB with CPU offloading
• License: MiniMax H3 Community License (with restrictions for certain regions)














Comments