Views
No views yet
hustvl/lightningdit-xl-imagenet256-800ep for local/offline use.LightningDit-XL-1-256pipeline.py (LightningDiTPipeline)transformer/transformer_lightningdit.py and weightsscheduler/scheduler_config.json (FlowMatchHeunDiscreteScheduler, shift=0.3)vae/ (REPA-E/vavae-hf)id2label in model_index.json, so class labels can be passed as ImageNet ids or English synonym strings.
LightningDiT-XL/1 at 256×256, 250 steps, CFG 6.7, cfg_interval_start=0.125, timestep_shift=0.3, seed 0.| Model | Resolution | Local path |
|---|---|---|
| LightningDiT-XL/1 | 256×256 | ./LightningDit-XL-1-256 |
1import torch
2from diffusers import DiffusionPipeline
3
4pipe = DiffusionPipeline.from_pretrained(
5 "./LightningDit-XL-1-256",
6 trust_remote_code=True,
7 torch_dtype=torch.bfloat16,
8).to("cuda")
9
10class_id = pipe.get_label_ids("golden retriever")[0]
11image = pipe(
12 class_labels=class_id,
13 num_inference_steps=250,
14 guidance_scale=6.7,
15 cfg_interval_start=0.125,
16 timestep_shift=0.3,
17 generator=torch.Generator(device="cuda").manual_seed(0),
18).images[0]1@inproceedings{yao2025reconstruction,
2 title={Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models},
3 author={Yao, Jingfeng and Yang, Bin and Wang, Xinggang},
4 booktitle={CVPR},
5 year={2025}
6}