BiliSakura/Looped-DiT-diffusers

Self-contained Looped-DiT 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.

Available checkpoints

Subfolder	Model	Params (denoiser + text encoder)	Patch	Loop depth	CFG
Looped-DiT-B-32/	Looped-DiT B/32	260M + 341M	32	4	6.0
Looped-DiT-B-16/	Looped-DiT B/16	258M + 341M	16	4	6.0

Benchmark scores (100 Euler steps, CFG 6.0, loop depth 4):

Model	GenEval	DPG-Bench	PRISM	CoReBench	SpatialGenEval	TIIF-Short	Avg
B/32 (290k)	85.1	85.3	54.4	44.5	52.3	76.1	66.3
B/16 (580k)	87.4	87.0	67.0	53.5	54.6	79.7	71.5

Repo layout

BiliSakura/Looped-DiT-diffusers/
├── README.md
├── .gitattributes
├── Looped-DiT-B-32/
│   ├── pipeline.py
│   ├── model_index.json
│   ├── demo.png
│   ├── scheduler/
│   ├── text_encoder/
│   ├── tokenizer/
│   └── transformer/
└── Looped-DiT-B-16/
    └── ...

Each variant is self-contained: load with custom_pipeline pointing at that folder’s pipeline.py and trust_remote_code=True. Looped-DiT denoises directly in RGB pixel space (no VAE).

Demo

Prompt: "a red cube on top of a blue sphere." — Looped-DiT B/16 at 512×512, 100 steps, guidance_scale=6.0, num_loops=4, torch_dtype=bfloat16, seed 42.

Same prompt and settings with Looped-DiT B/32.

Load from Hugging Face

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/Looped-DiT-diffusers",
    subfolder="Looped-DiT-B-16",
    custom_pipeline="pipeline.py",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
    "a red cube on top of a blue sphere",
    num_inference_steps=100,
    guidance_scale=6.0,
    num_loops=4,
    generator=generator,
).images[0]
image.save("demo.png")

For B/32, set subfolder="Looped-DiT-B-32".

Load from a local clone

from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("./Looped-DiT-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 red cube on top of a blue sphere",
    num_inference_steps=100,
    guidance_scale=6.0,
    num_loops=4,
    generator=generator,
).images[0]
image.save("demo.png")

Use ./Looped-DiT-B-32 instead of ./Looped-DiT-B-16 for the B/32 checkpoint.

Recommended inference settings

Variant	Resolution	Steps	CFG scale	num_loops	torch_dtype
Looped-DiT-B-32	512×512	100	6.0	4 (default)	bfloat16 (full pipeline)
Looped-DiT-B-16	512×512	100	6.0	4 (default)	bfloat16 (full pipeline)

Other loop depths work at inference when loop weights are shared (the default for released models).

Interface notes

• Text conditioning uses bundled google/flan-t5-large (T5EncoderModel + T5Tokenizer) in bfloat16, the same dtype as the denoiser. Prompt length is the tokenizer model_max_length (256).
• torch_dtype=torch.bfloat16 on from_pretrained sets both. Do not cast pipe.text_encoder back to float32.
• Set custom_pipeline to the variant’s pipeline.py (Hub: "pipeline.py" with subfolder; local: absolute path).
• 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.

Links

• Upstream B/32 weights: sensenova/Looped-DiT-B32
• Upstream B/16 weights: sensenova/Looped-DiT-B16
• Backbone: MiniT2I · BiliSakura/MiniT2I-diffusers

License

MIT (same as upstream Looped-DiT and MiniT2I).