Refactor app.py and download_qwen_image_models.py for improved model handling

Introduce a new image type handling mechanism in app.py and download_qwen_image_models.py, allowing for dynamic model selection based on the QWEN_IMAGE_TYPE environment variable. Update functions to generate filenames for both edit and layered models, enhancing flexibility in model management. Modify the training script to accommodate these changes, ensuring proper metadata generation and error handling for layered configurations. This refactor improves code clarity and user experience by streamlining model downloads and configurations.

app.py

@@ -49,21 +49,31 @@ def _bash_quote(s: str) -> str:
     return "'" + str(s).replace("'", "'\"'\"'") + "'"
 
 
-def _get_edit_model_version() -> str:
-    raw = os.environ.get("QWEN_IMAGE_EDIT_VERSION", "")
+_QWEN_IMAGE_TYPES = ("edit-2509", "edit-2511", "layered")
+
+
+def _get_qwen_image_type() -> str:
+    raw = os.environ.get("QWEN_IMAGE_TYPE", "edit-2509")
     v = raw.strip().lower()
-    if v in ("2511", "edit-2511", "edit_2511", "qwen_image_edit_2511", "qwen-image-edit-2511"):
-        return "edit-2511"
-    if v in ("2509", "edit-2509", "edit_2509", "qwen_image_edit_2509", "qwen-image-edit-2509"):
+    if v not in _QWEN_IMAGE_TYPES:
+        print(f"[QIE] Unsupported QWEN_IMAGE_TYPE: {v}. Defaulting to edit-2509.")
         return "edit-2509"
-    return "edit-2509"
+    return v
 
 
-def _edit_dit_filename(edit_version: str) -> str:
-    suffix = edit_version.split("-", 1)[1] if "-" in edit_version else edit_version
+def _dit_filename_for_type(image_type: str) -> str:
+    if image_type == "layered":
+        return "qwen_image_layered_bf16.safetensors"
+    suffix = image_type.split("-", 1)[1]
     return f"qwen_image_edit_{suffix}_bf16.safetensors"
 
 
+def _vae_filename_for_type(image_type: str) -> str:
+    if image_type == "layered":
+        return "qwen_image_layered_vae.safetensors"
+    return "diffusion_pytorch_model.safetensors"
+
+
 def _ensure_workspace_auto_files() -> None:
     """Ensure /workspace/auto has required helper files from this repo.
 
@@ -101,6 +111,7 @@ def _update_dataset_toml(
     train_batch_size: Optional[int] = None,
     control_res_w: Optional[int] = None,
     control_res_h: Optional[int] = None,
+    multiple_target: Optional[bool] = None,
 ) -> None:
     """Update dataset TOML for resolution/batch/control resolution in-place.
 
@@ -138,16 +149,22 @@ def _update_dataset_toml(
             block = _set_in_general(block, "batch_size", f"batch_size = {int(train_batch_size)}")
         txt = head + block + tail
 
-    if control_res_w and control_res_h:
-        m2 = re.search(r"(?ms)^\[\[datasets\]\]\s*(.*?)(?=^\[\[|\Z)", txt)
-        if m2:
-            head, block, tail = txt[:m2.start(1)], m2.group(1), txt[m2.end(1):]
+    m2 = re.search(r"(?ms)^\[\[datasets\]\]\s*(.*?)(?=^\[\[|\Z)", txt)
+    if m2:
+        head, block, tail = txt[:m2.start(1)], m2.group(1), txt[m2.end(1):]
+        if control_res_w and control_res_h:
             line = f"qwen_image_edit_control_resolution = [{int(control_res_w)}, {int(control_res_h)}]"
             if re.search(r"(?m)^\s*qwen_image_edit_control_resolution\s*=", block):
                 block = re.sub(r"(?m)^\s*qwen_image_edit_control_resolution\s*=.*$", line, block)
             else:
                 block = block.rstrip() + "\n" + line + "\n"
-            txt = head + block + tail
+        if multiple_target is not None:
+            mt_line = f"multiple_target = {'true' if multiple_target else 'false'}"
+            if re.search(r"(?m)^\s*multiple_target\s*=", block):
+                block = re.sub(r"(?m)^\s*multiple_target\s*=.*$", mt_line, block)
+            else:
+                block = block.rstrip() + "\n" + mt_line + "\n"
+        txt = head + block + tail
 
     try:
         Path(path).write_text(txt, encoding="utf-8")
@@ -244,6 +261,103 @@ def _norm_key(filename: str, prefix: str, suffix: str) -> str:
     return stem
 
 
+IMAGE_EXTENSIONS = (".jpg", ".jpeg", ".png", ".bmp", ".gif")
+
+
+def _list_image_files(folder: str) -> List[str]:
+    try:
+        files = [
+            fn
+            for fn in sorted(os.listdir(folder))
+            if fn.lower().endswith(IMAGE_EXTENSIONS)
+            and os.path.isfile(os.path.join(folder, fn))
+        ]
+        return files
+    except Exception:
+        return []
+
+
+def _join_posix(base: str, name: str) -> str:
+    if not base:
+        return name.replace("\\", "/")
+    return base.rstrip("/\\") + "/" + name.replace("\\", "/").lstrip("/")
+
+
+def _generate_layered_jsonl(
+    image_dir: str,
+    caption: str,
+    output_json: str,
+    control_dirs: List[Optional[str]],
+    target_prefix: str = "",
+    target_suffix: str = "",
+    control_prefixes: Optional[List[Optional[str]]] = None,
+    control_suffixes: Optional[List[Optional[str]]] = None,
+    allow_single: bool = True,
+) -> int:
+    base_files = _list_image_files(image_dir)
+    if not base_files:
+        raise ValueError("IMAGE_FOLDER に画像が見つかりません。")
+
+    active_controls = [d for d in control_dirs if d]
+    if not active_controls:
+        raise ValueError("Layered では少なくとも1つのレイヤー画像が必要です。")
+
+    use_name_matching = bool(target_prefix or target_suffix)
+    if control_prefixes and any((c or "").strip() for c in control_prefixes):
+        use_name_matching = True
+    if control_suffixes and any((c or "").strip() for c in control_suffixes):
+        use_name_matching = True
+
+    if not use_name_matching:
+        missing: List[str] = []
+        for fname in base_files:
+            for cdir in active_controls:
+                expected = os.path.join(cdir, fname)
+                if not os.path.exists(expected):
+                    missing.append(expected)
+        if missing:
+            preview = "\n".join(f" - {p}" for p in missing[:5])
+            raise ValueError(f"対応するレイヤーが見つかりません（strict）:\n{preview}")
+
+    out_dir = os.path.dirname(output_json)
+    if out_dir:
+        os.makedirs(out_dir, exist_ok=True)
+
+    count = 0
+    with open(output_json, "w", encoding="utf-8") as f:
+        for base_fn in base_files:
+            entry: Dict[str, str] = {
+                "image_path_0": _join_posix(image_dir, base_fn),
+                "caption": caption,
+            }
+            key = _norm_key(base_fn, target_prefix, target_suffix)
+            layer_index = 1
+            for i, cdir in enumerate(control_dirs):
+                if not cdir:
+                    continue
+                cprefix = ""
+                csuffix = ""
+                if control_prefixes and i < len(control_prefixes) and control_prefixes[i]:
+                    cprefix = control_prefixes[i]
+                if control_suffixes and i < len(control_suffixes) and control_suffixes[i]:
+                    csuffix = control_suffixes[i]
+                expected_name = base_fn if not use_name_matching else f"{cprefix}{key}{csuffix}.png"
+                expected_path = os.path.join(cdir, expected_name)
+                if not os.path.exists(expected_path):
+                    cfiles = _list_image_files(cdir)
+                    if allow_single and len(cfiles) == 1:
+                        expected_name = cfiles[0]
+                    else:
+                        raise ValueError(
+                            f"対応するレイヤーが見つかりません: expected={expected_name}, layer_index={i}"
+                        )
+                entry[f"image_path_{layer_index}"] = _join_posix(cdir, expected_name)
+                layer_index += 1
+            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
+            count += 1
+    return count
+
+
 def _copy_uploads(
     uploads: List[Tuple[str, str]], dest_dir: str, rename_to: Optional[List[str]] = None
 ) -> List[str]:
@@ -451,9 +565,10 @@ def _prepare_script(
             txt,
         )
 
-    edit_version = _get_edit_model_version()
-    dit_filename = _edit_dit_filename(edit_version)
-    txt = _replace_model_path(txt, "vae", "vae/diffusion_pytorch_model.safetensors")
+    image_type = _get_qwen_image_type()
+    dit_filename = _dit_filename_for_type(image_type)
+    vae_filename = _vae_filename_for_type(image_type)
+    txt = _replace_model_path(txt, "vae", f"vae/{vae_filename}")
     txt = _replace_model_path(txt, "text_encoder", "text_encoder/qwen_2.5_vl_7b.safetensors")
     txt = _replace_model_path(txt, "dit", f"dit/{dit_filename}")
 
@@ -554,7 +669,7 @@ def _prepare_script(
         _set_var('MAX_TRAIN_EPOCHS', str(override_max_epochs))
     if override_save_every is not None and override_save_every > 0:
         _set_var('SAVE_EVERY_N_EPOCHS', str(override_save_every))
-    _set_var('MODEL_VERSION', edit_version)
+    _set_var('MODEL_VERSION', image_type)
 
     # Write to a temp file alongside this repo for easier inspection
     run_dir = TRAINING_DIR / ".gradio_runs"
@@ -729,6 +844,10 @@ def run_training(
         yield (log_buf, ckpts, artifacts, run_out_dir)
         return
 
+    image_type = _get_qwen_image_type()
+    log_buf += f"[QIE] Model type: {image_type}\n"
+    yield (log_buf, ckpts, artifacts, run_out_dir)
+
     # Ensure /auto holds helper files expected by the script
     _ensure_workspace_auto_files()
     # Resolve data root and create dataset directories (auto-decide)
@@ -764,11 +883,17 @@ def run_training(
         _extract_paths(control6_uploads),
         _extract_paths(control7_uploads),
     ]
-    # Require control_0; others optional
-    if not control_upload_sets[0]:
-        log_buf += "[ERROR] control_0 images are required.\n"
-        yield (log_buf, ckpts, artifacts, run_out_dir)
-        return
+    if image_type == "layered":
+        if not any(control_upload_sets):
+            log_buf += "[ERROR] Layered では control_0〜7 のいずれかが必要です。\n"
+            yield (log_buf, ckpts, artifacts, run_out_dir)
+            return
+    else:
+        # Require control_0; others optional
+        if not control_upload_sets[0]:
+            log_buf += "[ERROR] control_0 images are required.\n"
+            yield (log_buf, ckpts, artifacts, run_out_dir)
+            return
 
     control_dirs: List[Optional[str]] = []
     for i, uploads in enumerate(control_upload_sets):
@@ -784,7 +909,10 @@ def run_training(
         log_buf += f"[QIE] Copied {len(uploads)} control_{i} images to {cdir}\n"
         yield (log_buf, ckpts, artifacts, run_out_dir)
 
-    # Metadata.jsonl will be generated by create_image_caption_json.py in train_QIE.sh
+    control_dirs_abs = [
+        (os.path.join(ds_dir, name) if name else None)
+        for name in control_dirs
+    ]
 
     # Prepare script with user parameters
     control_folders = [
@@ -805,15 +933,24 @@ def run_training(
 
     # Update dataset config with requested resolution/batch settings
     try:
+        ctrl_w = int(control_res_w) if control_res_w else None
+        ctrl_h = int(control_res_h) if control_res_h else None
+        if image_type == "layered":
+            ctrl_w = None
+            ctrl_h = None
         _update_dataset_toml(
             ds_conf,
             img_res_w=int(train_res_w) if train_res_w else None,
             img_res_h=int(train_res_h) if train_res_h else None,
             train_batch_size=int(train_batch_size) if train_batch_size else None,
-            control_res_w=int(control_res_w) if control_res_w else None,
-            control_res_h=int(control_res_h) if control_res_h else None,
+            control_res_w=ctrl_w,
+            control_res_h=ctrl_h,
+            multiple_target=(image_type == "layered"),
         )
-        log_buf += f"[QIE] Updated dataset config: resolution=({train_res_w},{train_res_h}), batch_size={train_batch_size}, control_res=({control_res_w},{control_res_h})\n"
+        if image_type == "layered":
+            log_buf += f"[QIE] Updated dataset config: resolution=({train_res_w},{train_res_h}), batch_size={train_batch_size}, multiple_target=true\n"
+        else:
+            log_buf += f"[QIE] Updated dataset config: resolution=({train_res_w},{train_res_h}), batch_size={train_batch_size}, control_res=({control_res_w},{control_res_h})\n"
     except Exception as e:
         log_buf += f"[QIE] WARN: failed to update dataset config: {e}\n"
     # Expose dataset config for download (if exists)
@@ -867,8 +1004,39 @@ def run_training(
         pass
     yield (log_buf, ckpts, artifacts, run_out_dir)
 
+    output_json = os.path.join(out_base, "metadata.jsonl")
+    if image_type == "layered":
+        _sync_dataset_config_jsonl(ds_conf, output_json)
+        log_buf += f"[QIE] Generating layered metadata: {output_json}\n"
+        yield (log_buf, ckpts, artifacts, run_out_dir)
+        try:
+            count = _generate_layered_jsonl(
+                image_dir=img_dir,
+                caption=caption,
+                output_json=output_json,
+                control_dirs=control_dirs_abs,
+                target_prefix=(target_prefix or ""),
+                target_suffix=(target_suffix or ""),
+                control_prefixes=control_prefixes,
+                control_suffixes=control_suffixes,
+                allow_single=True,
+            )
+            log_buf += f"[QIE] Layered metadata written: {count} entries\n"
+        except Exception as e:
+            log_buf += f"[ERROR] Layered metadata generation failed: {e}\n"
+            yield (log_buf, ckpts, artifacts, run_out_dir)
+            return
+
+        if os.path.isfile(output_json) and output_json not in artifacts:
+            artifacts.append(output_json)
+        yield (log_buf, ckpts, artifacts, run_out_dir)
+
+        if config_only:
+            log_buf += "[QIE] Config-only mode: skipping cache/training.\n"
+            yield (log_buf, ckpts, artifacts, run_out_dir)
+            return
+
     if config_only:
-        output_json = os.path.join(out_base, "metadata.jsonl")
         _sync_dataset_config_jsonl(ds_conf, output_json)
         log_buf += f"[QIE] Generating metadata: {output_json}\n"
         yield (log_buf, ckpts, artifacts, run_out_dir)
@@ -1258,6 +1426,10 @@ def _startup_download_models() -> None:
         os.makedirs(MODELS_ROOT_RUNTIME, exist_ok=True)
 
     print(f"[QIE] Ensuring models in: {MODELS_ROOT_RUNTIME}")
+    skip_raw = os.environ.get("QWEN_IMAGE_SKIP_DOWNLOAD", "")
+    if skip_raw.strip().lower() in ("1", "true", "yes", "on"):
+        print("[QIE] QWEN_IMAGE_SKIP_DOWNLOAD=1: skipping model download.")
+        return
     try:
         download_all_models(MODELS_ROOT_RUNTIME)
     except Exception as e:

download_qwen_image_models.py

@@ -6,23 +6,22 @@ from huggingface_hub import hf_hub_download
 # Enable hf_transfer for faster downloads
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
-DEFAULT_MODELS_DIR = os.environ.get("QWEN_IMAGE_MODELS_DIR", 
+DEFAULT_MODELS_DIR = os.environ.get("QWEN_IMAGE_MODELS_DIR",
                                     "/Qwen-Image_models")
 # Temporary download root (requested: use /tmp instead of /workspace)
 TMP_DOWNLOAD_ROOT = os.environ.get("QWEN_IMAGE_TMP_DIR", "/tmp/qie_downloads")
 
 
-def _normalize_edit_version(raw: str) -> str:
-    v = (raw or "").strip().lower()
-    if v in ("2511", "edit-2511", "edit_2511", "qwen_image_edit_2511", "qwen-image-edit-2511"):
-        return "edit-2511"
-    if v in ("2509", "edit-2509", "edit_2509", "qwen_image_edit_2509", "qwen-image-edit-2509"):
-        return "edit-2509"
-    return "edit-2509"
+_QWEN_IMAGE_TYPES = ("edit-2509", "edit-2511", "layered")
 
 
-def _get_edit_version() -> str:
-    return _normalize_edit_version(os.environ.get("QWEN_IMAGE_EDIT_VERSION", ""))
+def _get_qwen_image_type() -> str:
+    raw = os.environ.get("QWEN_IMAGE_TYPE", "edit-2509")
+    v = raw.strip().lower()
+    if v not in _QWEN_IMAGE_TYPES:
+        print(f"[QIE] Unsupported QWEN_IMAGE_TYPE: {v}. Defaulting to edit-2509.")
+        return "edit-2509"
+    return v
 
 
 def _ensure_dirs(root: str) -> None:
@@ -71,45 +70,64 @@ def _download_then_place(*, repo_id: str, filename: str, subfolder: str,
 
 
 def download_all_models(models_dir: str = DEFAULT_MODELS_DIR) -> Dict[str, str]:
-    """Download required Qwen-Image-Edit models into models_dir.
+    """Download required Qwen-Image models into models_dir.
 
     Returns a dict of component -> local file path.
     """
     _ensure_dirs(models_dir)
 
     print(f"[QIE] Models directory: {models_dir}")
+    image_type = _get_qwen_image_type()
+    print(f"[QIE] Model type: {image_type}")
 
     print("[QIE] Download dir (tmp):", TMP_DOWNLOAD_ROOT)
     print("[QIE] Final models dir:", models_dir)
 
-    edit_version = _get_edit_version()
-    dit_suffix = edit_version.split("-", 1)[1] if "-" in edit_version else edit_version
-    dit_filename = f"qwen_image_edit_{dit_suffix}_bf16.safetensors"
-
-    # Download to /tmp then move to final path
-    dit_path = _download_then_place(
-        repo_id="Comfy-Org/Qwen-Image-Edit_ComfyUI",
-        filename=dit_filename,
-        subfolder="split_files/diffusion_models",
-        component="dit",
-        models_dir=models_dir,
-    )
-
-    print("[QIE] Downloading VAE model(s)...")
-    vae_main = _download_then_place(
-        repo_id="Qwen/Qwen-Image-Edit",
-        filename="diffusion_pytorch_model.safetensors",
-        subfolder="vae",
-        component="vae",
-        models_dir=models_dir,
-    )
-    vae_alt = _download_then_place(
-        repo_id="Comfy-Org/Qwen-Image_ComfyUI",
-        filename="qwen_image_vae.safetensors",
-        subfolder="split_files/vae",
-        component="vae",
-        models_dir=models_dir,
-    )
+    if image_type == "layered":
+        dit_filename = "qwen_image_layered_bf16.safetensors"
+        dit_path = _download_then_place(
+            repo_id="Comfy-Org/Qwen-Image-Layered_ComfyUI",
+            filename=dit_filename,
+            subfolder="split_files/diffusion_models",
+            component="dit",
+            models_dir=models_dir,
+        )
+        print("[QIE] Downloading VAE model(s)...")
+        vae_main = _download_then_place(
+            repo_id="Comfy-Org/Qwen-Image-Layered_ComfyUI",
+            filename="qwen_image_layered_vae.safetensors",
+            subfolder="split_files/vae",
+            component="vae",
+            models_dir=models_dir,
+        )
+        vae_alt = ""
+    else:
+        edit_suffix = image_type.split("-", 1)[1]
+        dit_filename = f"qwen_image_edit_{edit_suffix}_bf16.safetensors"
+        # Download to /tmp then move to final path
+        dit_path = _download_then_place(
+            repo_id="Comfy-Org/Qwen-Image-Edit_ComfyUI",
+            filename=dit_filename,
+            subfolder="split_files/diffusion_models",
+            component="dit",
+            models_dir=models_dir,
+        )
+
+        print("[QIE] Downloading VAE model(s)...")
+        vae_main = _download_then_place(
+            repo_id="Qwen/Qwen-Image-Edit",
+            filename="diffusion_pytorch_model.safetensors",
+            subfolder="vae",
+            component="vae",
+            models_dir=models_dir,
+        )
+        vae_alt = _download_then_place(
+            repo_id="Comfy-Org/Qwen-Image_ComfyUI",
+            filename="qwen_image_vae.safetensors",
+            subfolder="split_files/vae",
+            component="vae",
+            models_dir=models_dir,
+        )
 
     print("[QIE] Downloading Text Encoder...")
     te_path = _download_then_place(

train_QIE.sh

@@ -3,7 +3,7 @@
 # ==============================
 # Generate metadata.jsonl before training
 # Configure variables directly in this file.
-# Environment variable override: QWEN_IMAGE_EDIT_VERSION (edit-2509 or edit-2511)
+# Environment variable override: QWEN_IMAGE_TYPE (edit-2509, edit-2511, layered)
 # ==============================
 
 echo "[QIE] Torch version check"
@@ -46,14 +46,20 @@ SEED=42
 MAX_TRAIN_EPOCHS=100
 SAVE_EVERY_N_EPOCHS=10
 
-# Model version (edit-2509 or edit-2511)
-MODEL_VERSION="${QWEN_IMAGE_EDIT_VERSION:-edit-2509}"
-if [[ "$MODEL_VERSION" != "edit-2509" && "$MODEL_VERSION" != "edit-2511" ]]; then
-  echo "[QIE] Unsupported QWEN_IMAGE_EDIT_VERSION: $MODEL_VERSION, defaulting to edit-2509"
+# Model type (edit-2509, edit-2511, layered)
+MODEL_VERSION="${QWEN_IMAGE_TYPE:-edit-2509}"
+if [[ "$MODEL_VERSION" != "edit-2509" && "$MODEL_VERSION" != "edit-2511" && "$MODEL_VERSION" != "layered" ]]; then
+  echo "[QIE] Unsupported QWEN_IMAGE_TYPE: $MODEL_VERSION, defaulting to edit-2509"
   MODEL_VERSION="edit-2509"
 fi
-DIT_SUFFIX="${MODEL_VERSION#edit-}"
-DIT_FILENAME="qwen_image_edit_${DIT_SUFFIX}_bf16.safetensors"
+if [[ "$MODEL_VERSION" == "layered" ]]; then
+  DIT_FILENAME="qwen_image_layered_bf16.safetensors"
+  VAE_FILENAME="qwen_image_layered_vae.safetensors"
+else
+  DIT_SUFFIX="${MODEL_VERSION#edit-}"
+  DIT_FILENAME="qwen_image_edit_${DIT_SUFFIX}_bf16.safetensors"
+  VAE_FILENAME="diffusion_pytorch_model.safetensors"
+fi
 
 # Build control args from folder names with auto-detect fallback
 CONTROL_ARGS=()
@@ -103,19 +109,23 @@ fi
 
 cd /workspace/auto
 
-echo "[QIE] Generating metadata: $OUTPUT_JSON"
-python create_image_caption_json.py \
-  -i "${DATASET_DIR%/}/${IMAGE_FOLDER}" \
-  -c "$CAPTION" \
-  -o "$OUTPUT_JSON" \
-  --image-dir "${DATASET_DIR%/}/${IMAGE_FOLDER}" \
-  "${CONTROL_ARGS[@]}"
+if [[ "$MODEL_VERSION" == "layered" ]]; then
+  echo "[QIE] Layered mode: skip metadata generation (expects prebuilt JSONL)."
+else
+  echo "[QIE] Generating metadata: $OUTPUT_JSON"
+  python create_image_caption_json.py \
+    -i "${DATASET_DIR%/}/${IMAGE_FOLDER}" \
+    -c "$CAPTION" \
+    -o "$OUTPUT_JSON" \
+    --image-dir "${DATASET_DIR%/}/${IMAGE_FOLDER}" \
+    "${CONTROL_ARGS[@]}"
+fi
 
 cd /musubi-tuner
 
 python qwen_image_cache_latents.py  \
 --dataset_config "$DATASET_CONFIG" \
---vae "/workspace/Qwen-Image_models/vae/diffusion_pytorch_model.safetensors" \
+--vae "/workspace/Qwen-Image_models/vae/${VAE_FILENAME}" \
 --model_version "$MODEL_VERSION" \
 --vae_spatial_tile_sample_min_size 16384
 
@@ -128,7 +138,7 @@ python qwen_image_cache_text_encoder_outputs.py  \
 accelerate launch src/musubi_tuner/qwen_image_train_network.py \
 --model_version "$MODEL_VERSION" \
 --dit "/workspace/Qwen-Image_models/dit/${DIT_FILENAME}" \
---vae "/workspace/Qwen-Image_models/vae/diffusion_pytorch_model.safetensors" \
+--vae "/workspace/Qwen-Image_models/vae/${VAE_FILENAME}" \
 --text_encoder "/workspace/Qwen-Image_models/text_encoder/qwen_2.5_vl_7b.safetensors" \
 --dataset_config "$DATASET_CONFIG"  \
 --mixed_precision bf16 \