JLT-diffusers

Native diffusers checkpoints for JLT (Clean-Latent Prediction in Latent Diffusion Transformers). Paper: arXiv:2605.27102. Code: akatsuki-neo/JLT. Each variant folder is self-contained:

• pipeline.py — JLTPipeline
• scheduler/scheduling_jlt.py — JLTScheduler (JLT time t: 0 → 1, Heun / Euler)
• transformer/transformer_jlt.py — JLTTransformer2DModel
• vae/ — bundled AutoencoderKLFlux2

Load with PyPI diffusers and trust_remote_code=True. No extra JLT package is required.

Demo

Class: golden retriever (ImageNet 207) — JLT-B/1 at 256×256, 50 Heun steps, guidance_scale=2.9, noise_scale=1.0, torch_dtype=bfloat16, seed 42. These checkpoints are trained and sampled at 256×256 only.

Available checkpoints

Checkpoint	Path	Resolution	Latent	Recommended CFG
JLT-B/1	JLT-B-1/	256×256	FLUX.2, patch /1	2.9
JLT-L/1	JLT-L-1/	256×256	FLUX.2, patch /1	2.9
JLT-H/1	JLT-H-1/	256×256	FLUX.2, patch /1	2.9

ImageNet class labels

Each variant keeps an English id2label map in model_index.json.

• pipe.id2label — id → English label
• pipe.labels — synonym → id
• pipe.get_label_ids("golden retriever")
• pipe(class_labels="golden retriever", ...) — string labels resolve automatically

Inference

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/JLT-diffusers",
    subfolder="JLT-B-1",  # or "JLT-L-1" / "JLT-H-1"
    custom_pipeline="pipeline.py",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

print(pipe.id2label[207])
image = pipe(
    class_labels="golden retriever",
    num_inference_steps=50,
    guidance_scale=2.9,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("demo.png")

Citation

@article{fu2026jlt,
  title={{JLT}: {C}lean-{L}atent {P}rediction in {L}atent {D}iffusion {T}ransformers},
  author={Fu, Funing and Wang, Tenghui and Zhou, Guanyu and Cen, Junyong and Zhu, Qichao},
  journal = {arXiv preprint arXiv:2605.27102},
  year={2026}
}