| name | echo-memory-eval |
| description | Run Echo-Memory replay, in-domain loop/revisit, and open-domain eval v2 scripts. Use when checking HF checkpoints, editing eval/v2, or tracing CKPT → memory profile via env/memory_baseline_runtime.py. |
Echo-Memory evaluation
Required env
export WAN_BASE_MODEL=/path/to/Wan2.1-T2V-1.3B
export DATASET_BASE_PATH=data/Context-as-Memory-Dataset
export PYTHONPATH=$PWD:${PYTHONPATH:-}
export CKPT=./ckpts/context_k1/epoch-0.safetensors
CKPT path must include the row folder name (e.g. context_k1/, context_k20/) so env/memory_baseline_runtime.py selects the right memory profile.
Main scripts
| Goal | Script |
|---|
| GT replay check (PSNR/SSIM/LPIPS) | eval/v2/run_basic_replay_gt.sh |
| Loop + in-domain + open-domain revisit | eval/v2/run_static_consistency_loop_and_revisit.sh |
| Revisit suite assets | eval/v2/revisit_suite/ |
Download a baseline:
huggingface-cli download Echo-Team/Echo-Memory context_k1/epoch-0.safetensors --local-dir ./ckpts
Index: doc/checkpoints.md · HF repo: Echo-Team/Echo-Memory
Eval branches (name these in prompts)
- Replay — short-horizon pixel fidelity
- In-domain — 180° loop / held layouts
- Open-domain — edited first frames, return probes (
assets/opendomain_revisit/)
Agent checklist
- Ask which branch (replay / in-domain / open-domain) unless obvious from context.
- Verify
CKPT row folder matches the memory family under test.
- Prefer extending existing
eval/v2/*.sh over duplicating inference loops.
- Optional vars:
EVALS_ROOT, NUM_SAMPLES_LOOP, MULTIVIEW_FIRSTFRAME_DIR — see eval/v2/README.md.