Skip to content

feat: open-source Sana-Video 2.0 - #439

Merged
lawrence-cj merged 8 commits into
mainfrom
release/sana-video-2.0
Aug 27, 2026
Merged

lawrence-cj merged 8 commits into
mainfrom
release/sana-video-2.0

Conversation

@lawrence-cj

@lawrence-cj lawrence-cj commented Aug 2, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

  • add the production SANA-Video 2.0 architecture with public 5B and 14B variants
  • reuse the shared train_video_scripts/train_video_ivjoint_chunk.py and inference_video_scripts/inference_sana_video.py entry points with minimal SANA-Video 2.0 compatibility
  • add clean 480p and 720p reference configs for the 5B variant; the 14B config and checkpoint are intentionally not included yet
  • release the jointly post-trained T2V/TI2V SANA-Video 2.0 5B 720p checkpoint
  • add a verified 5B 720p release demo generated from the public hf:// checkpoint, together with its exact prompt and launch command
  • add the SANA-Video 2.0 guide, model-zoo entry, README/news links, legacy asset-doc mirror, and focused tests
  • keep the public surface limited to release artifacts without internal experiments, cluster settings, private-data paths, or redundant version-specific runners

Architecture

  • 5B: 32 layers, hidden size 2560, 4 Attention Residual blocks, 4.47B parameters
  • 14B: 40 layers, hidden size 4096, 5 Attention Residual blocks, 14.25B parameters
  • both variants use the final hybrid gated linear/softmax attention design and shared timestep-independent Attention Residual blocks

Checkpoint

  • HF repo: Efficient-Large-Model/SANA-Video_2.0_5B_720p
  • path: checkpoints/SANA_Video_2.0_5B_720p.pth
  • SHA256: 7e557554540b4cbbc515166b43a7d307285ab250cedfb48627878227a722d25a
  • contains only the merged model state_dict (797 tensors); optimizer, scheduler, training state, and standalone LoRA tensors are excluded
  • the model card, matching release config, Apache-2.0 license, SANA-Video collection entry, and verified demo are published

Verified 5B 720p demo

SANA-Video 2.0 5B release demo: cartoon rooster in a floral vintage room

▶ Watch or download the generated MP4

  • output: H.264, 1280 × 736, 193 frames, 24 FPS, 8.04 seconds
  • video SHA256: e99a8e48ace067ac3f582e9ffeb9b0a9f42af7f0f8c0cec2d74054a5088397ce
  • assets commit: 82b9f1da

Prompt: In a cozy, vintage room adorned with floral wallpaper, a cartoon
rooster sits comfortably in a floral-patterned armchair, sipping from a bottle
of beer. The rooster, with its vibrant red comb and wattle, displays a range of
expressions—smiling, nodding, and opening its beak wide in a cheerful manner.
The setting includes wooden furniture and another beer bottle on the table,
adding to the relaxed atmosphere. The camera captures the rooster from a
close-up angle, emphasizing its animated movements and lively demeanor.

Exact command used inside the one-H100 allocation (identical in the root README, release docs, and model card):

bash inference_video_scripts/inference_sana_video.sh \
  --np 1 \
  --config configs/sana_video2/SanaVideo2_5B_720p.yaml \
  --model_path hf://Efficient-Large-Model/SANA-Video_2.0_5B_720p/checkpoints/SANA_Video_2.0_5B_720p.pth \
  --txt_file=asset/samples/sana_video2_5b_720p_demo.txt \
  --cfg_scale 8 \
  --flow_shift 12 \
  --step 50 \
  --fps 24 \
  --motion_score 20 \
  --seed 0 \
  --work_dir output/sana_video2_t2v_720p_demo

Impact

This exposes the model architecture, training and inference paths, release documentation, and the first SANA-Video 2.0 checkpoint. The released 5B model supports 720p, approximately 8-second text-to-video and text-image-to-video generation. The 14B architecture remains public, while its config and checkpoint remain unreleased.

Validation

  • strict merge verification: all 256 LoRA target weights match the merge formula and every non-LoRA tensor remains exact
  • release checkpoint verification: exact key order/coverage, tensor metadata/value equality, single state_dict container, and local/remote SHA256 match
  • H100 BF16 smoke test: strict 797-tensor load and finite forward output passed (8.41 GiB peak allocated)
  • end-to-end H100 inference from the public hf:// checkpoint passed with the exact 50-step command above; sampling, VAE decode, and MP4 encoding completed successfully
  • generated-video verification: 193 decoded frames, 24 FPS, 1280 × 736, 8.04 seconds, exact local/remote asset SHA256 match, and visual inspection of the first/middle/final frames passed
  • released LTX 2.3 VAE compatibility was fixed for Diffusers 0.38.0; the 726,105,969-parameter VAE then loaded completely in BF16
  • pytest -q tests/test_sana_video2.py — 10 passed
  • both shared entry points passed their --help startup checks
  • MkDocs build passed; warnings remain limited to seven pre-existing issues outside the SANA-Video 2.0 page
  • full pre-commit run --all-files passed
  • privacy, cluster-path, secret-pattern, stale-reference, and diff-integrity scans passed

@lawrence-cj
lawrence-cj force-pushed the release/sana-video-2.0 branch from bccf6ac to db48e16 Compare August 2, 2026 07:12
@lawrence-cj lawrence-cj self-assigned this Aug 2, 2026
@lawrence-cj
lawrence-cj marked this pull request as ready for review August 20, 2026 11:23
@lawrence-cj lawrence-cj added the run-tests Run CI Test label Aug 22, 2026
@github-actions github-actions Bot removed the run-tests Run CI Test label Aug 22, 2026
@lawrence-cj
lawrence-cj requested a review from HeliosZhao August 22, 2026 06:24

@HeliosZhao HeliosZhao left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@lawrence-cj
lawrence-cj merged commit 220c7fe into main Aug 27, 2026
12 checks passed
@lawrence-cj
lawrence-cj deleted the release/sana-video-2.0 branch August 27, 2026 05:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants