Ruihan-org/NeuralCarving-checkpoints

NeuralCarving research checkpoints

The selected October 2026 checkpoint backup is complete: 63,777,174,329 bytes of weights and training state, committed and verified by SHA-256. Configurations, normalization and selection provenance are included separately.

Directory	Contents and purpose
normal-vae-v2/ordinary-step003000/	Selected fine-tuned normal VAE encoder and decoder, safetensors and architecture JSON.
dit-v2/ordinary-step020000.pt	Ordinary DiT weights used in the October 27 evaluation, paired with the selected V2 VAE.
dit-v2/resume-step020000/	Matching EMA and optimizer/training state; ordinary weights are the file above.
dit-v2/resume-step021000/	Latest October training checkpoint: ordinary weights, EMA weights and optimizer/training state.
normal-vae-v1/resume-step003000/	V1 full training checkpoint.
normal-vae-v2/resume-step003000/	V2 full training checkpoint.
normal-vae-v3/resume-step003000/	V3 full training checkpoint.

The evaluated ordinary step-20,000 DiT and ordinary step-3,000 V2 VAE are the selected inference pair. The step-21,000 files are preserved as the latest resume state; they are not a claim of a better evaluated model. V1/V3 are alternative experimental endpoints, not replacements for the selected V2 pair.

Normal-VAE targets represent world-space D−N perturbations, distinct from absolute-normal conditioning images. The included configuration, normalization, hashes and source identity describe the required research implementation. This repository does not provide a standalone inference library, and archival verification is not a new training-resume test.

Model code is derived from TRELLIS.2. Preserve upstream licensing and model requirements.