---
type: "post"
id: "it-edits-with-up-to-8-reference-images-4bb0b9"
slug: "it-edits-with-up-to-8-reference-images-4bb0b9"
title: "It edits with up to 8 reference images and runs natively in ComfyUI and Diffusers."
description: "PrunaAI’s P-Image-OSS is an image-generation and editing model based on MiniMax-H3, with eight-step 8-megapixel generation and editing from as many as eight images."
retrieval_nugget: "PrunaAI’s P-Image-OSS is an image-generation and editing model based on MiniMax-H3, with eight-step 8-megapixel generation and editing from as many as eight images. Its Diffusers pipeline and separate ComfyUI files make it usable in different workflows, but GPU requirements and a geographically restricted license are important deployment constraints."
published_at: "2026-11-20"
updated_at: "2026-11-20"
record_date: "2026-11-20"
date_kind: "published_at"
topics: ["generative-media","multimodal","models"]
entities: ["PrunaAI","P-Image-OSS","MiniMax-H3","Diffusers","ComfyUI","Qwen3.5-9B","MiniMax"]
related_patterns: ["hardware-adoption-is-routed-through-familiar-ai-software-stacks","generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://huggingface.co/PrunaAI/P-Image-OSS"
source_urls: ["https://huggingface.co/PrunaAI/P-Image-OSS"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:it-edits-with-up-to-8-reference-images-4bb0b9"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/it-edits-with-up-to-8-reference-images-4bb0b9/","markdown":"https://newruntime.com/posts/it-edits-with-up-to-8-reference-images-4bb0b9.md","json":"https://newruntime.com/posts/it-edits-with-up-to-8-reference-images-4bb0b9.json"}
markdown_url: "https://newruntime.com/posts/it-edits-with-up-to-8-reference-images-4bb0b9.md"
---

# It edits with up to 8 reference images and runs natively in ComfyUI and Diffusers.

## Retrieval answer

PrunaAI’s P-Image-OSS is an image-generation and editing model based on MiniMax-H3, with eight-step 8-megapixel generation and editing from as many as eight images. Its Diffusers pipeline and separate ComfyUI files make it usable in different workflows, but GPU requirements and a geographically restricted license are important deployment constraints.

## Summary
PrunaAI’s P-Image-OSS is an image-generation and editing model based on MiniMax-H3, with eight-step 8-megapixel generation and editing from as many as eight images. Its Diffusers pipeline and separate ComfyUI files make it usable in different workflows, but GPU requirements and a geographically restricted license are important deployment constraints.

## What happened
PrunaAI’s Hugging Face model card describes P-Image-OSS as a text-to-image and image-editing model based on MiniMax-H3. The model uses a 33.1-billion-parameter transformer, a 4-billion-parameter text encoder, a VAE, and a prompt upsampler. It generates images in eight steps, without classifier-free guidance, a negative prompt, or a guidance scale. Text-to-image output can reach 8 megapixels; edits work best at 4 or 8 megapixels. The card says the model works best at 8 megapixels, while lower resolutions are more prone to artifacts in faces, hands, and small details.

Editing accepts one image or a list of up to eight. The first image is the one to edit; the others become references that the prompt can identify as Picture 2, Picture 3, and so on. The included prompt upsampler, based on Qwen3.5-9B, is enabled by default. For edits, it turns a short request into a prompt describing what to change and what to preserve. The model card says a request such as “make the pickup truck red” may cause the model to redraw the scene, while an edit prompt that states what should remain unchanged can limit the change. Users can also remove the upsampler from the pipeline and write that more detailed prompt themselves.

The repository provides a Diffusers pipeline and example code for generation and editing. The card says 8-megapixel inference in bf16 needs about 101 GB of GPU memory with the prompt upsampler, or 83 GB without it. With automatic CPU offloading, it reports a peak of about 76 GB on an 80-GB GPU such as an H100. Its B200 performance figures use batch size one, bf16, 8-megapixel output, and no compilation or offloading: text-to-image takes about nine seconds without the upsampler, which adds six to eleven seconds. For ComfyUI, PrunaAI points users to a separate repository with int8 and 4-bit files that the model card says run in 21 GB of VRAM.

PrunaAI says the model is strongest at photorealistic, film, and editorial looks, and graphic styles with short quoted text. It warns that long prompts and hype words can worsen results; hands may break, small background figures can distort, and long or dense text is unreliable. The model produces still images, not video or audio. P-Image-OSS uses the MiniMax H3 Community License. The card says that license excludes the European Union, United Kingdom, Republic of Korea, and United States unless the user has a separate MiniMax license, and requires products built on the model to display “Powered by MiniMax H3.” The text-encoder components and prompt upsampler have separate Apache 2.0 terms.

## What the archive adds
New Runtime’s [earlier coverage of Pruna AI’s P-H3-Image](https://newruntime.com/posts/pruna-ai-open-sources-p-h3-image-an-76af40/) on 2026-11-03 described a separate MiniMax-H3-based image model with eight-step generation, 8-megapixel output, and editing with up to eight images. P-Image-OSS is a new release, not a restatement: it adds a documented Diffusers pipeline, ComfyUI quantized files, and detailed memory, performance, and usage guidance.

## Trend signal
This does not yet fit a pattern New Runtime tracks.

## Why it matters
A single editing interface for a base image and multiple references can make image composition more practical: teams can ask for a change while pointing to separate visual sources, instead of rebuilding every detail in a long prompt. The prompt upsampler is consequential because its job is to articulate both the requested edit and what should remain intact. That is useful workflow support, but it does not remove the need to inspect outputs, especially for text, hands, and small details.

The cost of that workflow is substantial hardware demand. The published memory figures imply that a team must plan for high-memory GPUs, offloading, or quantized ComfyUI weights rather than assuming a standard local development machine will run the full bf16 pipeline. The license also makes geography and product labeling part of the deployment decision, not paperwork to defer until launch. Teams evaluating the model should test it against their own image types and review license eligibility before integrating it into a production workflow.

## What happens next
- **PrunaAI will document additional memory or deployment configurations for P-Image-OSS** (6 months, 55%). Watch for: new repository or model-card guidance on supported GPUs and quantized pipelines.

## Limits and uncertainty
The speed and memory figures are PrunaAI’s published measurements, not an independent benchmark; the card specifies a B200 setup but does not establish performance across other hardware, workloads, or batch sizes. It also does not quantify output quality or compare it with other editing models through a controlled evaluation. The 21-GB figure applies to separate ComfyUI quantized files, not the bf16 Diffusers pipeline. Users should verify the current license terms and their eligibility directly before deployment. The card’s stated strengths and failure modes are vendor descriptions, and teams still need to test whether results meet their own requirements.

## Skill to build now
Build a small edit-evaluation set before adopting P-Image-OSS: choose representative source images, include reference images where your workflow needs them, and record both the short request and the generated edit prompt. Check whether requested changes happen while faces, layout, lighting, and other protected details stay intact. Include examples with small text and hands, which the model card identifies as weak spots. Run that test on the hardware and license terms you can actually use.

## Reader takeaway
P-Image-OSS combines multi-reference editing with a prompt-upsampled Diffusers workflow and ComfyUI quantized files. It is worth evaluating when image editing is your use case, but the full bf16 path demands high GPU memory, the vendor’s performance claims are not independent benchmarks, and the MiniMax H3 license sharply limits where it can be used without separate permission.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Generative Media](https://newruntime.com/topics/generative-media.md): Explore the generative-media topic hub.
- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
