OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Echo Memory

Echo Memory is research code for comparing Context, Compression, Spatial and State-Space memory in action-conditioned video world models. It includes a shared Wan 2.1 1.3B backbone, training recipes, data utilities, released checkpoints, and replay and revisit evaluation assets.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Echo Memory — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it for research or prototyping that needs an explicit comparison of video-memory strategies, especially when evaluating scene consistency across camera movement and revisits.

Who it is for

  • Researchers studying memory mechanisms in generative video world models.
  • Developers prototyping memory modules for chunk-wise, action-conditioned video generation.
  • Practitioners evaluating whether generated scenes retain identity, layout and viewpoint after camera-driven revisits.

Documented uses

  • Compare four memory families under a shared video backbone.
  • Generate video from a prompt and checkpoint, with optional first-frame conditioning and camera-action control.
  • Run static replay and revisit evaluations using released assets.
  • Inspect how scene identity, layout and viewpoint change when a generated trajectory returns to an earlier location.
Before you choose

Requirements & limits

  • The documented local stack uses Python and PyTorch on Linux.
  • The README lists an approximately 340 GB static in-domain evaluation pool, plus the Wan base model and checkpoint, for its end-to-end evaluation workflow.
  • The project provides a hosted demo described as generating a five-second clip from a first frame, prompt and camera action.
  • Dynamic SpatialVID training and inference recipes are included, while dynamic evaluation beyond static replay and revisit is marked as unfinished.
  • The repository is released under CC-BY-4.0.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run a released checkpoint on a single video

Generate one video from a released Echo Memory checkpoint using the unified inference script.

Before you start

  • A Linux environment with the repository's Python and PyTorch dependencies installed.
  • The Wan 2.1 T2V 1.3B base model and a released Echo Memory checkpoint.
  • A prompt; a first-frame image and camera-action file can also be supplied.

Steps

  1. Install the environment using the repository's environment.yml and requirements.txt instructions.

  2. Set WAN_BASE_MODEL to the Wan 2.1 base-model directory and add the repository to PYTHONPATH.

  3. Run inference/unified_inference.py with the checkpoint, prompt and output path; optionally provide --context_image and --action_path.

Expected outcome

The inference script is configured to save a generated video at the chosen output path.

Guide limitations

  • This is a documented workflow, not an execution test.
  • The guide covers single-chunk generation; it does not establish that a particular machine has sufficient compute or that every checkpoint will work in every environment.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
Echo-Team-Joy-Future-Academy-JD/Echo-Memory · Official homepage title · retained in its source language
How we review sourcesSuggest a correction