doubixz/primitive-operation-painter-2-1b

Primitive Operation Painter 2 — 1B

POP 2 / Primitive Operation Painter v3.3：供原生 Primitive Painter 应用使用的 1B BF16 推理权重。

本模型根据中文提示词、已有图元，或两者结合，自回归地逐个预测绘画操作。输出是可编辑的矩形、椭圆等原生图元序列，而不是直接输出像素。此发布为当前应用使用的 Epoch 60 / global step 10020，包含配套的原版 SentencePiece 分词器。

This second-generation Primitive Operation Painter is a custom autoregressive drawing model with Chinese text conditioning and editable primitive-operation outputs. It is the v3.3 Epoch 60 checkpoint used by the native Windows application, not a generic language model or a diffusion model.

下载与使用 / Download and use

下载以下两个文件，在原生应用的“设置”中选择对应路径：

文件	用途
model.pt	原生程序直接读取的 BF16 权重，约 2.01 GB / 1.87 GiB
caption_spm_1500.model	原版 1500 词表 SentencePiece 分词器

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

Select these files in a compatible native Primitive Painter application's settings. The checkpoint tensor layout matches native app 0.9.27. This repository does not install the application or GPU runtime.

The exact original model.pt is preserved, without quantization or conversion. BF16 tensors are under model_state_dict. The file includes training metadata and an optimizer-format descriptor, not the optimizer shards needed for exact training resumption. This is an inference release, not a full resumable training checkpoint.

config.json documents this custom architecture. It does not make the model compatible with transformers.AutoModel or standard text-generation inference servers; use the matching Primitive Painter implementation.

架构 / Architecture

Item	Value
Parameters	1,004,182,914
Transformer layers	35
Hidden width	1536
Attention heads / head width	24 / 64
Gated FFN hidden width	4096
Weight precision	BF16
Total context	1024 text/primitive positions
Text vocabulary	1500
Primitive/control vocabulary	7069
Attribute output vocabulary	7054
Native canvas	512 × 512
Positional encoding	RoPE

Each primitive has nine discrete attributes: x, y, angle, width, height, shape, red, green, blue. Complete primitive-history records combine nine attribute embeddings into one model position. Partial predictions use learned empty-attribute embeddings and shared-history branches. Text tokens occupy separate positions before the primitive BOS.

The drawing protocol includes global, coarse-grid and fine-grid stages, with explicit region-control, permanent-end and interruption tokens. The application supports canvas editing, continuation and branching history. Its context-overclocking feature selects retained history; it does not increase the 1024-position attention window.

生成 / Generation

支持文字输入、图元前缀输入及两者结合。可从 CFG=1、Rényi α=4 开始试用，再按画面效果调整。这是采样设置示例，不是质量保证。

Classifier-free guidance compares conditional and empty-text predictions; CFG=1 is the conditional baseline. Rényi adaptive candidate selection, batching, region scheduling and optional KV-cache precision are runtime features. They do not alter the distributed BF16 weights.

来源 / Checkpoint provenance

• Project generation: POP 2; internal project version: v3.3.
• Completed epoch 60, global optimizer step 10020.
• EMA disabled; these are the checkpoint's model weights, not the previous model's EMA release.
• Recorded training configuration: 3 ranks, per-device micro-batch 10, gradient accumulation 20, effective global batch 600.
• checkpoint_metadata.json records architecture, token layout and training metadata, excluding model tensors and optimizer state.

限制 / Limitations

Developed around anime-face drawing data and Chinese descriptions; this is not a general-purpose photorealistic image generator. Prompt following and visual quality vary. The native application may reject out-of-vocabulary text. Outputs are constrained to the primitive vocabulary and region protocol. GPU support and performance depend on application build, runtime and hardware; this release makes no universal speed or compatibility claim.

Training images, prompt datasets, optimizer shards, user projects, application binaries and GPU runtimes are not included. Users must ensure they have the necessary rights to their input data and intended uses.

项目与旧版 / Project and earlier release

• Code project: wonderfulearth/primitive-operation-painter. Previously published code targets the earlier model; use the updated native implementation for this checkpoint.
• Earlier POP 1: doubixz/primitive-operation-painter-weight, a different 0.3B / 144-step model. Its architecture, token layout and loader are not interchangeable with this release.

License

MIT, consistent with the project's existing public release. See LICENSE. This does not redistribute or grant rights to excluded third-party training data.