Apppublic

hugging-apps/sana-video-2-0-5b-720p-4step

sourceHugging Faceapache-2.0updated 16h agoView on Hugging Face

0likes

App README

SANA-Video 2.0 · 5B · 720p · 4-Step Preview

Project page · 4-step model · 50-step base model · Documentation · GitHub · Paper

This Space serves the independent four-step distilled `SANA-Video 2.0 5B 720p preview` checkpoint. It generates 736×1280 videos with four sampling steps, using a five-second (81 frames at 16 FPS) or eight-second (193 frames at 24 FPS) temporal profile. The Space uses the latest global-step-1000 tri-temporal DMD EMA export and exposes duration, motion conditioning, and RL LoRA strength controls.

Exact preview settings

Setting	Value
Denoising stages	4
Resolution	736×1280
Source-tower construction size	480 (dynamic RoPE; explicit 736×1280 bucket)
Duration profiles	5 seconds: 81 frames / 16 FPS; 8 seconds: 193 frames / 24 FPS
CFG	1.0
Motion-score suffix	adjustable from 0 to 100; default `20` (`0` disables it)
RL LoRA scale	adjustable from `0.3` to `1.0` in `0.1` increments; default `0.7`
Negative prompt	empty
Default demo seed	`4`
Generator sigma profile	`sana_shift6_dpm`
Physical sigmas	`0.9998332262, 0.9471688271, 0.8568977118, 0.6663702130`
Model timesteps	`999, 947, 856, 666`

Each non-terminal stage predicts x0 and re-noises it at the next fixed sigma with a fresh draw from the same seeded generator. flow_shift is not applied by this sampler. Resolution, CFG, and the four-step schedule stay fixed; the UI exposes the validated duration profiles, Motion Score, RL LoRA scale, and seed.

Pipeline

Component	Source
Transformer	`Efficient-Large-Model/SANA-Video_2.0_5B_720p_4step` (`checkpoints/SANA_Video_2.0_5B_720p_4step.pth`)
RL scale controller	Same repository (`adapters/SANA_Video_2.0_5B_720p_RL500_LoRA_rank128.pt`)
VAE	`Efficient-Large-Model/LTX-2.3-Diffusers` (`vae/`, 128-channel latents, 8×32×32 stride)
Text encoder	`Efficient-Large-Model/gemma-2-2b-it` (decoder only, 300-token window plus CHI prompt)

The transformer is the latest global-step-1000 full-model EMA export distilled from the ReFL step-500 initialization. It is not replaced by a LoRA and contains no optimizer, scheduler, or training-loop state. The separately published rank-128 RL LoRA is used only to adjust RL strength around that full model at runtime:

text

W(s) = W(scale=1) + (s - 1) × (alpha / rank) × (B_RL @ A_RL)

Scale 1.0 reproduces the uploaded full checkpoint. The default 0.7 setting is a post-hoc adjustment around that reference, not a separately DMD-trained model. Every scale change is reconstructed from immutable BF16 reference weights, so repeated slider changes do not accumulate rounding error.

Acknowledgements & attribution

The diffusion/ package in this Space is NVlabs/Sana (Apache-2.0), pruned to the modules needed for SANA-Video 2.0 inference. The inference path, the four-step DMD sampler (FastVideoDMD4Step), and the RL LoRA scale controller follow the authors' deployment-aligned rollout as exposed by the official `Efficient-Large-Model/sana-video2-5b-720p-demo` Space (Apache-2.0).

The example prompts are selected from the official SANA-Video 2.0 project page curated set, hosted in `Efficient-Large-Model/Sana-assets` (MIT-licensed).

The tiny mmcv/ package is a compatibility stand-in that provides only the four symbols touched by the vendored inference code (Registry, build_from_cfg, Config/ConfigDict, and get_dist_info).

The preview_controls.py input-contract helpers mirror the official release contract for the validated duration profiles and RL scale interval.