Primitive Operation Painter 2 1B

Primitive Operation Painter 2 1B

A 1B drawing model that predicts editable shapes such as rectangles and ellipses from a Chinese prompt, instead of pixels. Weights for the native Primitive Painter app on Windows, not for Transformers.

Made by doubixz on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Primitive Operation Painter 2 is a 1B drawing model, the second generation of the project. It does not output pixels. It predicts one drawing operation at a time, such as a rectangle or an ellipse, from a Chinese text prompt, from shapes already on the canvas, or from both. The result is an editable list of native shapes. This release holds the weights of the project's version v3.3, trained for 60 epochs, with the matching tokenizer. It is built for the native Windows Primitive Painter application.

Who it suits: It suits developers and tinkerers who use or build the native Primitive Painter app, and who want editable shape drawings from Chinese prompts.

Industry | Who uses it | How they use it | Example
Digital art | Hobbyist illustrators | Sketch editable shape drawings from a Chinese prompt | A prompt makes a list of ellipses and rectangles you can still move
Tool development | Developers of the native app | Load the checkpoint for inference | The card says the tensor layout matches app version 0.9.27
Research | Generative model researchers | Study drawing as a sequence of operations | Each shape has nine attributes, from position to color

You can skip it if you want photorealistic images or a Transformers model, since the card says this is neither a general image generator nor compatible with standard text-generation servers.

Where to use it:

• The native Primitive Painter application on Windows, through its settings
• A GPU machine, since the card says performance depends on the app build and hardware
• Python with the Hugging Face tools, to download the files
• Canvas editing, continuation and branching history, which the app supports

Use case | Who | What they get back
Prompt to drawing | Illustrators | An editable list of shapes on a 512 × 512 canvas
Continue a sketch | App users | New shapes added to the shapes already drawn
Branch a drawing | Experimenters | Several versions from the same history

How to set it up: You need the native application and the two files.

• A compatible native Primitive Painter application
• The model.pt file, about 2.01 GB of BF16 weights
• The caption_spm_1500.model tokenizer file
• Python with the Hugging Face tools, to download them

1. Install a compatible version of the native application. This repository does not install it or the GPU runtime.
2. Download the files with the command below.
3. Open the application's settings.
4. Select the two downloaded files.

hf download doubixz/primitive-operation-painter-2-1b --local-dir POP2-1B

How to use it: Type a Chinese prompt, give existing shapes, or give both. The card suggests starting with CFG at 1 and Rényi alpha at 4, then adjusting by eye. It calls these sample settings, not a quality guarantee.

Tool | What it does
model.pt | The BF16 weights the native app reads
caption_spm_1500.model | The 1500-word tokenizer for Chinese prompts
config.json | Documents the custom architecture, but does not make it work with AutoModel
checkpoint_metadata.json | Records the architecture, token layout and training details

Example: You type a Chinese prompt in the app, with CFG at 1. The model predicts shapes one after another, first a global stage, then coarse and fine grids. You then edit any shape by hand, since each is a native shape.

For newcomers and experts: Think of it as a painter that works with a few shape tools instead of a pixel brush. A first small try is one short Chinese prompt with the card's starting settings.

Each shape carries nine values: position, angle, size, shape type and three color channels. The nine attribute embeddings are combined into one model position. Drawing runs in three stages: a global pass, then coarse and fine grid passes, with special tokens for region control and ending. The context-overclocking feature selects which history to keep and does not widen the 1024-position window.

Type a Chinese prompt or add shapes
The model predicts a global layout
It refines in coarse and fine grids
Shapes appear as editable primitives
Edit, continue or branch the drawing

Good to know: The card lists these architecture figures.

Title: Architecture
Item | Value
Parameters | 1,004,182,914
Transformer layers | 35
Hidden width | 1536
Attention heads / head width | 24 / 64
Total context | 1024 positions
Text vocabulary | 1500
Native canvas | 512 × 512

• Training: the card records epoch 60 and global step 10020, with 3 ranks and an effective batch of 600
• Release type: inference only, with no optimizer shards for resuming training
• Language: Chinese
• License: MIT, without rights to the excluded third-party training data
• Cost to run: none named, but it needs the native app and a GPU
• Data it reaches: the files you download and the prompts and shapes you use locally
• Not for: photorealistic images, standard text-generation servers or work that needs a general model

On Hugging Face, 13 November 2026

• Task: text to image
• Downloads in the last 30 days: 119
• Likes: 4
• License: mit
• Languages: zh
• Tags: primitive_operation_painter_2, primitive-operation-painter, autoregressive, image-generation, text-to-image, vector-graphics, bf16

Links

• Model: https://huggingface.co/doubixz/primitive-operation-painter-2-1b
• Files: https://huggingface.co/doubixz/primitive-operation-painter-2-1b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download doubixz/primitive-operation-painter-2-1b

Then follow the usage in its model card. Nothing in this kit runs by itself.