MiniT2I Diffusers Checkpoints

This private repository contains the Diffusers-compatible PyTorch weights for both MiniT2I-B/16 and MiniT2I-L/16. MiniT2I-B/16 uses the JAX checkpoint EMA decay 0.99995, and MiniT2I-L/16 uses EMA decay 0.9999; both are exported from step 290K. Load one repository, then select the model at inference time with model_type.

Models

Aliases such as b, base, minit2i-b/16, l, large, and minit2i-l/16 are also supported.

Usage

import torch
from diffusers import DiffusionPipeline

HUB_MODEL_ID = "MiniT2I/MiniT2I"

pipe = DiffusionPipeline.from_pretrained(
    HUB_MODEL_ID,
    custom_pipeline=HUB_MODEL_ID,
    trust_remote_code=True,
)

image = pipe(
    "A lonely astronaut standing on a quiet beach under two moons.",
    model_type="b16",
    guidance_scale=2.5,
    num_inference_steps=100,
    torch_dtype=torch.bfloat16,
).images[0]
image.save("minit2i-b16.png")

image = pipe(
    "a watercolor painting of a mountain lake at sunrise",
    model_type="l16",
    guidance_scale=6.0,
    num_inference_steps=100,
    torch_dtype=torch.bfloat16,
).images[0]
image.save("minit2i-l16.png")

The selected submodel is downloaded lazily from this repository, so calling with model_type="b16" does not download the L/16 weights.

Links

• Blog: Text-to-Image Generation Made Simple
• PyTorch/Diffusers release: Hope7Happiness/t2i-release
• JAX release: PeppaKing8/minit2i-jax

Related Checkpoints

Original JAX checkpoints are stored separately in private repositories:

• MiniT2I/MiniT2I-B-16-jax for MiniT2I-B/16
• MiniT2I/MiniT2I-L-16-jax for MiniT2I-L/16

Citation

@misc{minit2i2026,
  title  = {MiniT2I: A Minimalist Baseline for Text-to-Image Synthesis},
  author = {Wang, Xianbang and Zhao, Hanhong and Lu, Yiyang and Zhou, Kangyang and Ma, Linrui and He, Kaiming},
  year   = {2026},
  url    = {https://peppaking8.github.io/#/post/minit2i}
}