SiT-IBOT checkpoints

Archive of existing SiT-IBOT ImageNet research checkpoints.

This repository archives 577 existing ImageNet experiment checkpoints (about 1.93 TB). All 577 checkpoint paths and byte sizes were verified after upload; see upload-verification.json. It includes intermediate and ablation checkpoints, not 577 independent final models. See checkpoint_manifest.csv for relative download paths and byte sizes, and checkpoint_inventory.csv for original experiment/launcher references.

Main checkpoints

• Baseline, 1M steps: checkpoints/SiT-XL-self-flow-ibot-no-class-reverse-remove/009-SelfFlowDiT-XL-2-Linear-velocity-None/1000000.pt (2,721,817,888 bytes).
• Method, 1M steps: checkpoints/SiT-XL-self-flow-with-ibot_reverse_remove_simple_head_cls_w0.2_clslayer18/008-SelfFlowDiT-XL-2-Linear-velocity-None/1000000.pt (2,743,066,840 bytes).

Both use SelfFlowDiT-XL/2, a CLS token, same-view patch alignment weight 0.8, and the matched two-view generative objective. The method adds cross-view CLS alignment at block 18 with weight 0.2; baseline CLS weight is zero. Use the ema entry for sampling and evaluation. The main 1M files have had the raw student and optimizer states removed; later full checkpoints preserve training continuation. The archive has mixed checkpoint contents, documented where pruning manifests are available.

Use

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    "Nicholas0228/SiT-IBOT",
    "checkpoints/SiT-XL-self-flow-with-ibot_reverse_remove_simple_head_cls_w0.2_clslayer18/008-SelfFlowDiT-XL-2-Linear-velocity-None/1000000.pt",
)

The source package is maintained at https://github.com/Nicholas0228/SiT-IBOT (access may be restricted). These are research PyTorch checkpoints containing Python metadata, rather than a Diffusers pipeline or Safetensors export. ImageNet/VOC data and the stabilityai/sd-vae-ft-ema VAE are external dependencies.