BiliSakura/MiniT2I-diffusers

Self-contained MiniT2I text-to-image checkpoints for Hugging Face diffusers. Each variant folder ships its own pipeline code, component modules, bundled FLAN-T5-Large text encoder, and transformer weights.

Converted from MiniT2I/MiniT2I using MiniT2I-diffusers in Visual-Generative-Foundation-Model-Collection.

Available checkpoints

Subfolder	Model	Params (denoiser + text encoder)	Patch	Recommended CFG
MiniT2I-B-16/	MiniT2I-B/16	258M + 341M	16	2.5
MiniT2I-L-16/	MiniT2I-L/16	912M + 341M	16	6.0

Repo layout

BiliSakura/MiniT2I-diffusers/
├── README.md
├── MiniT2I-B-16/
│   ├── pipeline.py
│   ├── model_index.json
│   ├── conversion_metadata.json
│   ├── demo.png
│   ├── scheduler/
│   │   └── scheduler_config.json
│   ├── text_encoder/
│   ├── tokenizer/
│   └── transformer/
│       ├── config.json
│       ├── diffusion_pytorch_model.safetensors
│       └── transformer_minit2i.py
└── MiniT2I-L-16/
    └── ...

Each variant is self-contained: load with custom_pipeline=.../pipeline.py and trust_remote_code=True. MiniT2I denoises directly in RGB pixel space (no VAE).

Demo

Prompt: "A lonely astronaut standing on a quiet beach under two moons." — MiniT2I-B/16 at 512×512, 100 steps, guidance_scale=2.5, seed 42.

Load from Hugging Face

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/MiniT2I-diffusers/MiniT2I-B-16",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
    "A lonely astronaut standing on a quiet beach under two moons.",
    num_inference_steps=100,
    guidance_scale=2.5,
    generator=generator,
).images[0]
image.save("demo.png")

For MiniT2I-L/16, use MiniT2I-L-16 and guidance_scale=6.0.

Load from a local clone

from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("./MiniT2I-B-16").resolve()
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    local_files_only=True,
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
    "A lonely astronaut standing on a quiet beach under two moons.",
    num_inference_steps=100,
    guidance_scale=2.5,
    generator=generator,
).images[0]
image.save("demo.png")

Load a variant subfolder (e.g. ./MiniT2I-B-16), not the repo root.

Recommended inference settings

Variant	Resolution	Steps	CFG scale	torch_dtype
MiniT2I-B-16	512×512	100	2.5	bfloat16
MiniT2I-L-16	512×512	100	6.0	bfloat16

For GenEval / DPG-Bench evaluation, upstream configs use guidance_scale=5.0 for both B/16 and L/16.

Interface notes

• Text conditioning uses bundled google/flan-t5-large (T5EncoderModel + T5Tokenizer).
• Scheduler is FlowMatchEulerDiscreteScheduler with 1000 training timesteps and shift=1.0.
• guidance_scale > 1.0 enables classifier-free guidance with an empty-string null prompt.
• Output resolution is fixed at 512×512 for these exports.

Regenerate bundles

From the repository root:

conda activate rsgen
python scripts/convert_minit2i_to_bilisakura.py

Links

• Blog: MiniT2I: A Minimalist Baseline for Text-to-Image Generation
• Upstream checkpoints: MiniT2I/MiniT2I
• PyTorch/Diffusers source: MiniT2I-diffusers

Citation

@misc{minit2i2026,
  title  = {MiniT2I: A Minimalist Baseline for Text-to-Image Generation},
  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}
}