#!/usr/bin/env bash # C3 baseline (gradient-based, labelled as such): official ESD (rohitgandikota/erasing @ pinned commit, MIT), # unmodified, repo defaults; output converted to our candidate format. # usage: scripts/run_esd.sh set -euo pipefail VARIANT=$1; MODEL_ID=$2; CONCEPT_TEXT=$3; METHOD=$4; KEY=$5 COMMIT=2ff19a10d6084fababbdcb75e10deb56bd21fddb if [ ! -d third_party/esd ]; then git clone -q https://github.com/rohitgandikota/erasing.git third_party/esd git -C third_party/esd checkout -q $COMMIT fi OUT=runs/edits/c3-esd-$VARIANT-$KEY mkdir -p $OUT/raw # ESD loads with use_safetensors=True; export the pinned revision as a local safetensors snapshot # (tiny-sd ships only .bin), which also pins ESD to exactly the weights every other condition uses. LOCAL=runs/base-snapshots/$VARIANT if [ ! -f $LOCAL/model_index.json ]; then python - "$VARIANT" "$LOCAL" <<'PY' import sys from diffusers import StableDiffusionPipeline from uvw.models import load_registry v = load_registry()[sys.argv[1]] pipe = StableDiffusionPipeline.from_pretrained(v.base_repo, revision=v.base_revision, safety_checker=None, requires_safety_checker=False) pipe.save_pretrained(sys.argv[2], safe_serialization=True) print("exported", v.base_repo, v.base_revision, "->", sys.argv[2]) PY fi MODEL_ID=$(realpath $LOCAL) START=$(date +%s) (cd third_party/esd && python esd_sd.py --basemodel_id "$MODEL_ID" --erase_concept "$CONCEPT_TEXT" \ --train_method "$METHOD" --save_path "../../$OUT/raw") echo "{\"train_seconds\": $(( $(date +%s) - START ))}" > $OUT/raw/timing.json python - "$OUT" "$METHOD" <<'PY' import json, sys from pathlib import Path import torch from safetensors.torch import load_file, save_file from uvw.provenance import tensor_digest out, method = Path(sys.argv[1]), sys.argv[2] files = sorted((out / "raw").rglob("*.safetensors")) + sorted((out / "raw").rglob("*.pt")) assert files, "no ESD checkpoint found" f = files[0] sd = load_file(str(f)) if f.suffix == ".safetensors" else torch.load(f, map_location="cpu") sd = {(k if k.startswith("unet.") else f"unet.{k}"): v.to(torch.float32).contiguous() for k, v in sd.items()} d = out / "candidates" / f"esd-{method}" d.mkdir(parents=True, exist_ok=True) save_file(sd, str(d / "edited_tensors.safetensors")) (d / "meta.json").write_text(json.dumps({"name": d.name, "condition": "C3", "method": f"ESD {method} (official, gradient-based)", "commit": "2ff19a10d6084fababbdcb75e10deb56bd21fddb", "source_file": f.name, "timing": json.loads((out / "raw" / "timing.json").read_text()), "tensors": {k: tensor_digest(v) for k, v in sd.items()}})) print(d.name, len(sd), "tensors from", f.name) PY