Inko

A 0.5B text-to-image model for fine art, in English and Chinese.

GitHub & technical report · Try the demo

Inko is a text-to-image model trained for painting and illustration across Chinese and Western traditions. Its 532M-parameter flow-matching transformer pairs a curated four-domain art corpus with caption and resolution curricula, chunked Muon optimization on Ascend, and DiffusionART post-training against a self-produced prompt and rubric pool. The full engineering record — data, pretraining, post-training and evaluation — is in the GitHub repository.

Release model: the EMA of the specialized post-training run (60 DiffusionART rounds on the completed 300-round main run), built on the 600k pretrained EMA. The pretrained base remains available at tag pretrain-base.

Gallery

All images are generated by the release model at 20 Euler steps and CFG 4.5. Prompts, seeds and canvas sizes are in the manifests (hero1, hero2).

Results

The chart compares nine models on five public benchmarks (GenEval, DPG, OneIG English, TIIF testmini and Qwen-Image-Bench), each run under its native inference recipe: Inko, the pretrained Inko base, and seven external references. Inko competes directly with the similarly sized DreamLite, SANA-0.6B and PixArt-Σ, and approaches the much larger SD3.5 Medium on DPG, OneIG alignment and Qwen-Image-Bench. Dense prompt alignment is its strongest general result, and its diversity is high relative to models with comparable alignment. Protocol, full tables, per-benchmark analysis and a fine-art comparison suite: the evaluation chapter.

Quickstart

git clone https://github.com/Mtrya/inko
cd inko
pip install -e .

import torch
from src.pipeline import InkoPipeline

pipe = InkoPipeline.from_pretrained(
    "kaupane/Inko", torch_dtype=torch.bfloat16
).to("cuda")

image = pipe(
    "中国水墨画，远山笼罩在薄雾中，近处一棵松树伸向江面，一叶小舟停在岸边。",
    width=1920,
    height=1088,
    generator=torch.Generator(device="cpu").manual_seed(42),
).images[0]
image.save("inko.png")

The release defaults are 896 × 896, 20 Euler steps, and CFG 4.5. Width and height must be multiples of 16, up to 1920 per axis; above the 1184-pixel training reference, Dy-YaRN scaling applies automatically. Use Python 3.12 and a CUDA GPU.

Model at a glance

Dimension	Choice
Model	Hidden width 1152, 16 heads, 1 double-stream + 24 single-stream blocks, FFN width 3072, 532,496,992 parameters
Text	Frozen Qwen3-0.6B, encoded online with an early exit after layer 20
Latents	Qwen-Image VAE, 16 channels, factor 8; 2×2 latent patches
Objective	Rectified flow with logit-normal timestep sampling and resolution time shift
Resolution	Variable-aspect 256p → 640p → 896p; 75:20:5 split of 600,000 optimizer updates
Post-training	DiffusionART with attention LoRA (rank 64), Muon 1e-3, EMA 0.999; seven reward channels dominated by five rubric categories
Training hardware	Ascend (16×910B2C) for both pretraining and post-training

Training journey

• Data: four domains (Chinese painting, Western art, people and anatomy, world knowledge), about 1.7M eligible rows at 256p, bilingual 50:50 captions, with per-dataset galleries and the curation story of each source.
• Pretraining: architecture under the 0.6B budget, packing and the measured CUDA/Ascend kernel work (+33.92% throughput), the loss-spike investigation, and the full-run stability record.
• Post-training: the DiffusionART selection campaign, the 24.6k-row rollout pool (kaupane/inko-rollout-25k), the rubric reward, and the two production runs.
• Evaluation: five benchmark catalogs, nine models, and a fixed fine-art comparison suite with per-case analysis.

Limitations

• Exact rendered text and spatial relations remain weak.
• Anatomy still breaks at times; reasoning, counting and world knowledge remain weak.
• Negation in dense scenes is unreliable: asked for a busy intersection with no blue cars, Inko, like most references, paints the cars.
• The corpus includes research-only sources; per-sample license metadata is preserved in the training configuration.