| DiTGS camera-fixed Residual Cross-Attention checkpoint |
| Copyright 2026 DiTGS contributors. |
|
|
| This checkpoint incorporates pretrained and fine-tuned components originating from: |
|
|
| TokenGS, Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. |
| Source: https: |
| Weights: https: |
| Verified snapshot: a7eb20340b99abd29b554c29c339446aebccff36 |
| Original base SHA256: ac6cb611e3b19ea4aaaeae32c952cf7555f68ef3e70b03199f24c0860245bba8 |
| The upstream root README states Apache License 2.0. An additional model card at |
| this same snapshot names the NVIDIA Internal Scientific Research and Development |
| Model License. Both declarations are disclosed in UPSTREAM_NOTICES.md. |
| This release relies on the upstream root release declaration for TokenGS; |
| it does not represent that the conflicting model-card text has been removed. |
|
|
| FlashWorld: Xinyang Li, Tengfei Wang, Zixiao Gu, Shengchuan Zhang, |
| Chunchao Guo, Liujuan Cao. |
| Code: https: |
| Pretrained weights: https: |
| Revision: 896f3964301821d0bdd92ccb9560fbf51b855599 (CC BY-NC-SA 4.0). |
| DiTGS retains and fine-tunes the latent-velocity transformer; FlashWorld's |
| Gaussian-feature output channels and reconstruction decoder are not distributed. |
|
|
| Wan2.2: Wan-AI team. |
| https: |
| Revision: b8fff7315c768468a5333511427288870b2e9635 (Apache-2.0). |
|
|
| Changes: camera-corrected multiview preprocessing/conditioning, feature adapter, |
| independent DiT-memory attention and KV projections, per-channel residual scaling, |
| and fine-tuning the Wan backbone and TokenGS Gaussian decoder/head with the |
| 10,000-update warmup plus 20,000-update joint DiTGS schedule. |
|
|
| The pretrained Wan VAE and TokenGS image encoder are retained frozen. |
| No optimizer, EMA, RNG state, dataset images or external text encoder is included. |
| Upstream names are attribution only and do not imply endorsement. |
| The source-code Apache-2.0 license is separate from the checkpoint license. |
| All applicable upstream notices and terms remain in effect. |
|
|
