Add Echo-Memory optional loader and Wan 1.3B example

[WayneJin0918]

Summary

• Add an optional Echo-Memory loader that overlays the released context_k1 DiT fine-tune onto official Wan-AI/Wan2.1-T2V-1.3B.
• Matching keys use original DiffSynth / Wan names, so no conversion is required. Extra action-MLP / SSM / spatial slots are skipped.
• Example + low-VRAM example, plus a short EN/中文 note in the Wan docs.

This is a community fine-tune overlay, not a new backbone. The multi-chunk camera-action / revisit protocol stays in Echo-Memory (arXiv:2606.09803).

Test plan

• python examples/wanvideo/model_inference/Echo-Memory.py with Wan 2.1 1.3B and Hugging Face access to Echo-Team/Echo-Memory
• Confirm the loader prints ~825 overlaid DiT keys and skipped action/SSM extras

Made with Cursor

[Artiprocher]

@WayneJin0918 Thanks for your contribution! We'll merge this PR after some code cleanup.

Also, for future development, feel free to check out our skills guide. It can help you write code that aligns with the framework's best practices more quickly: https://modelscope.cn/collections/DiffSynth-Studio/DiffSynth-Studio-Model-Integration-Skills

[WayneJin0918]

Thanks @Artiprocher — appreciated. We'll leave the cleanup to you unless you want a specific change from our side.

If it helps, the loader is intentionally a thin overlay (pipe.dit.load_state_dict(..., strict=False)), not a new backbone. Extra action / SSM keys are skipped on purpose.

Skills guide noted for later rows.

[Artiprocher]

@WayneJin0918 As this open-source project is built upon the ModelScope platform, we would kindly ask you to upload your model to ModelScope first. Once that's done, we'll be more than happy to provide full support for it. Thank you for your understanding!

[Artiprocher]

@WayneJin0918 Additionally, we noticed that the model you provided doesn't include the complete set of weights. Would you mind packaging it into a full weight file? This would allow us to integrate the model in line with the framework's best practices and ensure proper support for training. We really appreciate your help!

[WayneJin0918]

Thanks @Artiprocher — we'll do both.

1. ModelScope: we'll upload a DiffSynth-facing context_k1 repo (same Apache-2.0 row as Echo-Team/Echo-Memory) and post the model_id here.

2. Full weights: the current HF file is a research overlay (epoch-0.safetensors) — matching Wan DiT keys plus extra action/SSM slots, loaded with strict=False on official Wan 2.1 1.3B. We'll publish a complete DiT as diffusion_pytorch_model.safetensors (official Wan 1.3B DiT keys with the Echo-Memory overlay already merged; extras dropped) so it loads like a normal Wan transformer. T5 / VAE stay the official Wan-AI/Wan2.1-T2V-1.3B files.

If you also want the research extras (action_mlp / SSM) in that file for training, we can add a second full checkpoint.

Happy to point #1612 at the ModelScope id once it's up.