Claude/update lam modelscope fz c6 l

UNFIX_license

@@ -0,0 +1,33 @@
+Copyright (c) 2026 UNFIX Entertainment, Inc. All rights reserved.
+
+PROPRIETARY AND CONFIDENTIAL
+
+This software, including all accompanying documentation, source code, object code, 
+and related materials (collectively, the "Software"), is the proprietary property 
+of UNFIX Entertainment, Inc. 
+
+Unauthorized copying, reproduction, modification, distribution, display, or use 
+of this Software, or any portion thereof, via any medium is strictly prohibited 
+without the express prior written permission of UNFIX Entertainment, Inc. 
+
+---
+THIRD-PARTY COMPONENT ATTRIBUTION & ACKNOWLEDGMENTS
+
+Portions of this Software are based on or incorporate modifications to software 
+licensed under the Apache License, Version 2.0 (the "Apache License"). 
+
+This Software also acknowledges the use of concepts and/or foundational code from 
+the "Large Avatar Model (LAM)", as detailed in the following academic publication:
+He, Y., Gu, X., Ye, X., Xu, C., Zhao, Z., Dong, Y., Yuan, W., Dong, Z., & Bo, L. (2025). 
+"LAM: Large Avatar Model for One-shot Animatable Gaussian Head." Proceedings of the 
+Special Interest Group on Computer Graphics and Interactive Techniques Conference.
+
+You may not use the Apache-licensed portions of the Software except in compliance 
+with the Apache License. You may obtain a copy of the Apache License at:
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed 
+under the Apache License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES 
+OR CONDITIONS OF ANY KIND, either express or implied. See the Apache License 
+for the specific language governing permissions and limitations under the License.

concierge_modal.py

@@ -0,0 +1,156 @@
+# Copyright (c) 2024-2025, The Alibaba 3DAIGC Team Authors.
+# Modal environment and pipeline initialization for LAM avatar generation.
+
+import os
+import modal
+
+MINUTES = 60
+
+# ---------------------------------------------------------------------------
+# Modal Volume: holds model weights, assets, and sample motions
+# ---------------------------------------------------------------------------
+vol = modal.Volume.from_name("lam-model-vol", create_if_missing=True)
+
+# ---------------------------------------------------------------------------
+# Modal Image: full build environment for LAM inference
+# ---------------------------------------------------------------------------
+image = (
+    modal.Image.from_registry(
+        "nvidia/cuda:12.1.1-cudnn8-devel-ubuntu22.04",
+        add_python="3.10",
+    )
+    # System deps: Blender, FBX SDK, GL libs, ffmpeg, zip
+    .apt_install(
+        "git", "wget", "unzip", "xz-utils", "libgl1", "libglib2.0-0",
+        "libsm6", "libxrender1", "libxext6", "ffmpeg", "zip",
+    )
+    # Blender 4.0.2 (for GLB conversion via generateARKITGLBWithBlender)
+    .run_commands(
+        "wget -q https://download.blender.org/release/Blender4.0/blender-4.0.2-linux-x64.tar.xz"
+        " -O /tmp/blender.tar.xz",
+        "tar xf /tmp/blender.tar.xz -C /opt/",
+        "ln -s /opt/blender-4.0.2-linux-x64/blender /usr/local/bin/blender",
+        "rm /tmp/blender.tar.xz",
+    )
+    # FBX SDK (required by generateARKITGLBWithBlender)
+    .run_commands(
+        "pip install --no-cache-dir fbx-sdk-py",
+    )
+    # Core Python deps
+    .pip_install(
+        "torch==2.5.1", "torchvision==0.20.1",
+        "--index-url", "https://download.pytorch.org/whl/cu121",
+    )
+    .pip_install(
+        "safetensors", "omegaconf", "trimesh", "Pillow",
+        "opencv-python-headless", "scipy", "einops",
+        "transformers", "accelerate", "huggingface_hub",
+        "moviepy==1.0.3", "imageio[ffmpeg]",
+        "chumpy", "numpy==1.23.0",
+        "kiui", "plyfile",
+    )
+    # pytorch3d (pre-built wheel for py310/cu121/pt251)
+    .pip_install(
+        "pytorch3d",
+        find_links="https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py310_cu121_pyt251/download.html",
+    )
+    # DINOv2 dependencies
+    .pip_install("xformers==0.0.29.post1")
+    # Copy the LAM source tree into the container
+    .copy_local_dir(".", "/root/LAM", ignore=[
+        ".git", "__pycache__", "*.pyc", "output", "exps",
+        "train_data", ".venv", "node_modules",
+    ])
+    .env({
+        "NUMBA_THREADING_LAYER": "forksafe",
+        "PYTHONPATH": "/root/LAM",
+        "BLENDER_PATH": "/usr/local/bin/blender",
+    })
+)
+
+# ---------------------------------------------------------------------------
+# Pipeline initialisation (called once per container warm-up)
+# ---------------------------------------------------------------------------
+
+def _init_lam_pipeline():
+    """
+    Initialise and return (cfg, lam, flametracking).
+
+    Mirrors the official app_lam.py launch_gradio_app() init sequence.
+    """
+    import torch
+    from omegaconf import OmegaConf
+    from safetensors.torch import load_file
+
+    os.chdir("/root/LAM")
+
+    # ---- env ----
+    os.environ.update({
+        "APP_ENABLED": "1",
+        "APP_MODEL_NAME": "./model_zoo/lam_models/releases/lam/lam-20k/step_045500/",
+        "APP_INFER": "./configs/inference/lam-20k-8gpu.yaml",
+        "APP_TYPE": "infer.lam",
+        "NUMBA_THREADING_LAYER": "forksafe",
+    })
+
+    # ---- parse config (simplified, no argparse) ----
+    cfg = OmegaConf.create()
+    cfg_train = OmegaConf.load("./configs/inference/lam-20k-8gpu.yaml")
+    cfg.source_size = cfg_train.dataset.source_image_res          # 512
+    cfg.render_size = cfg_train.dataset.render_image.high          # 512
+    cfg.src_head_size = getattr(cfg_train.dataset, "src_head_size", 112)
+    cfg.motion_video_read_fps = 30
+    cfg.blender_path = os.environ.get("BLENDER_PATH", "/usr/local/bin/blender")
+    cfg.model_name = os.environ["APP_MODEL_NAME"]
+
+    model_name = cfg.model_name
+    _relative_path = os.path.join(
+        cfg_train.experiment.parent,
+        cfg_train.experiment.child,
+        os.path.basename(model_name).split("_")[-1],
+    )
+    cfg.save_tmp_dump = os.path.join("exps", "save_tmp", _relative_path)
+    cfg.image_dump = os.path.join("exps", "images", _relative_path)
+    cfg.video_dump = os.path.join("exps", "videos", _relative_path)
+    cfg.model = cfg_train.model
+
+    # ---- build model ----
+    from lam.models import ModelLAM
+
+    model = ModelLAM(**cfg.model)
+    resume = os.path.join(model_name, "model.safetensors")
+    print("=" * 80)
+    print(f"Loading pretrained weights from: {resume}")
+    if resume.endswith("safetensors"):
+        ckpt = load_file(resume, device="cpu")
+    else:
+        ckpt = torch.load(resume, map_location="cpu")
+    state_dict = model.state_dict()
+    for k, v in ckpt.items():
+        if k in state_dict:
+            if state_dict[k].shape == v.shape:
+                state_dict[k].copy_(v)
+            else:
+                print(f"[WARN] shape mismatch {k}: ckpt {v.shape} vs model {state_dict[k].shape}")
+        else:
+            print(f"[WARN] unexpected key {k}: {v.shape}")
+    print(f"Finished loading weights from: {resume}")
+    print("=" * 80)
+
+    lam = model
+    lam.to("cuda")
+    lam.eval()
+
+    # ---- FLAME tracking ----
+    from tools.flame_tracking_single_image import FlameTrackingSingleImage
+
+    flametracking = FlameTrackingSingleImage(
+        output_dir="output/tracking",
+        alignment_model_path="./model_zoo/flame_tracking_models/68_keypoints_model.pkl",
+        vgghead_model_path="./model_zoo/flame_tracking_models/vgghead/vgg_heads_l.trcd",
+        human_matting_path="./model_zoo/flame_tracking_models/matting/stylematte_synth.pt",
+        facebox_model_path="./model_zoo/flame_tracking_models/FaceBoxesV2.pth",
+        detect_iris_landmarks=False,
+    )
+
+    return cfg, lam, flametracking