sana-camera-latest-states

ZiyangYe/sana-camera-latest-states

sana-camera-latest-states is a machine learning model from ZiyangYe. Use it for the machine learning task on the model card, and read the license before you ship it in a product.

This repository is public so it can be cloned from other clusters. It contains only the latest resumable checkpoints and state files for the Sana-CameraVideo Stage1/Stage2 training lanes. Historical milestone checkpoi…

Downloads · 30 days

Access

Public

Updated Jul 2, 2026

Repo size

138 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.bin104 GB · 75%

At a glance

Access: Public
Created: Jul 1, 2026
Updated: Jul 2, 2026
SHA: f37cc4b2

Model card

From the Hugging Face model README

Sana-CameraVideo Latest Training States

Code repo: https://github.com/Pollyannnna/Sana-CameraVideo

Required External Assets

These base assets are not included here and must be placed on the new cluster as described in the code repo README:

SANA-Video 2B 480p DiT: /data/share/lijiang/ckpt/multi-agent-wm/Sana/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth
WanVAE 480p: /data/share/lijiang/ckpt/multi-agent-wm/Sana/SANA-Video_2B_480p/vae/Wan2.1_VAE.pth
Worldmem-VPT raw data for Stage1: /data/share/Worldmem-VPT/training/vpt
Solaris duet data for Stage2: /data/share/lijiang/data/mutli-agent-wm/solaris/datasets/solaris-training-dataset/duet/train

Layout

Each latest-step* folder has:

full_state/: full Accelerate/FSDP resume state (metadata.pth, model/pytorch_model_fsdp.bin, model/optimizer.bin, model/scheduler.bin, random states). Use this for same-topology exact resume.
model_only/: model-only .pth sidecar. Use this for inference or cross-topology continuation with optimizer reset and train.start_global_step set to the step in the folder name.
run_state/: config, selected git history, W&B run id when available, and train log.
manifests/latest_status.json: code commits, source paths, and stage metadata.

Latest Stages

Stage	Step	Path	Resume note
Stage1 4-GPU no-prompt final lane	89500	`stage1/stage1-4g-final/latest-step89500`	Use `full_state/` for same 4-rank resume, or `model_only/epoch_32_step_89500.pth` for model-only continuation.
Stage1 16-GPU no-prompt continuation	37500	`stage1/stage1-16g-np/latest-step37500`	Use the one-token no-prompt text shim; do not resume with prompt/Gemma config.
Stage1 learned null seed lane	10500	`stage1/stage1-text-null/latest-step10500`	Clean Stage1 seed used by Solaris duet Stage2.
Stage2 Solaris duet	15000	`stage2/solaris-duet/latest-step15000`	Multiplayer Stage2 state initialized from Stage1 text-null.

Clone

git lfs install
git clone https://huggingface.co/<owner>/sana-camera-latest-states

For continuation on a new cluster, clone code separately and checkout the corresponding branch:

git clone https://github.com/Pollyannnna/Sana-CameraVideo.git
cd Sana-CameraVideo
git checkout stage2  # or stage1 / stage1-text-null as needed

Created: Jul 1, 2026
Updated: Jul 2, 2026