cyypoi/GroundingDiT-TMM-AB-checkpoints

GroundingDiT TMM: A/B experiment checkpoint archive

Powered by Stability AI.

Research and recovery archive for the Ref-AVS source-pool A/B revision experiments. The model repository is public; downloading its files requires an approved access request. This archive is not the original paper's main-experiment checkpoint and is not a ready-to-use Diffusers pipeline.

An upload is complete only when ARCHIVE_COMPLETE.json exists and all payloads in ARCHIVE_MANIFEST.json have been verified at its recorded revision. Files visible during an upload can be incomplete as a collection.

Experiments and contents

Prefix	Experiment	Available checkpoints
A/initial	Ref-AVS-only A, original eight-GPU run	Epoch 6 / step 5298
A/continued	A continuation archived from the two-GPU A800 environment	Epochs 7 through 30
B/initial	Derived mixed-task B v6, source-pool queries with a 2:1:1 task configuration	Epochs 1 through 6
B/continued	B continuation archived from the two-GPU A800 environment	Epochs 7 through 20
A/exports	Existing A epoch-6 full-model evaluation export	Epoch 6

Where present, each archived run also includes its existing evaluation export, training/evaluation logs, frozen source, configuration, checkpoint index and completion markers. RUN_INDEX.json records the original source-run names and checkpoint markers. Raw benchmark media, dataset caches, frozen backbone caches and authentication credentials are excluded.

The epoch-6 A/B checkpoints share an update-count endpoint of 5298. Batch size, data construction and schedule must be checked in each run's configuration; epoch labels or update counts alone do not establish identical data exposure or a causal task-diversity control. Later A/B endpoints differ (A: 30 epochs; B: 20 epochs); do not compare them as matched budgets. Checkpoint index paths can refer to their original cluster location. Resolve the same checkpoint-directory basename under this archive's corresponding prefix.

Recovery and integrity

Install huggingface_hub, authenticate with an approved account, then download a pinned commit:

python restore_archive.py --revision <verified-commit-SHA> --destination ./groundingdit_AB

To retrieve a smaller subset, repeat --prefix, for example:

python restore_archive.py --revision <verified-commit-SHA> --destination ./groundingdit_AB \
  --prefix A/initial/checkpoints/epoch_06_step_005298/ \
  --prefix B/initial/checkpoints/epoch_06_step_005298/

The recovery helper downloads and verifies SHA256 without deserializing checkpoint pickle files. Training shards are trusted PyTorch serialization containing model, optimizer and RNG state as written by their original trainer. Use the recorded PyTorch version, distributed world size, code and configuration when restoring; the eight-rank A checkpoint and two-rank continuation checkpoints are different layouts. File-level upload verification does not establish a successful distributed training resume. Dataset roots and cache paths in archived configurations must be remapped to your environment. Some continuation archives retain logs and configurations without a full source snapshot.

Licensing and provenance

The transformer was fine-tuned from Stable Diffusion 3.5 Medium for mask-latent prediction; an audio projection component and segmentation training states are included. The original base-model license is distributed in LICENSE.md; see NOTICE.txt. This archive does not grant access to benchmark data or redistribute the base model's text encoders or VAE. No new evaluation, restoration success or main-table equivalence is claimed by publishing these files.