Upload folder using huggingface_hub

README.md

@@ -1,3 +1,59 @@
----
-license: mit
----
+---
+license: mit
+base_model: microsoft/TRELLIS-image-large
+pipeline_tag: image-to-3d
+tags:
+  - image-to-3d
+  - 3d-generation
+  - scene-generation
+  - trellis
+---
+
+# IScene-v1
+
+IScene-v1 is the first public-release checkpoint package for IScene inference.
+
+This repository is intended to be used with the IScene code release. It contains the IScene-specific checkpoint files and minimal configuration needed to run single-image, segmentation-conditioned 3D scene generation.
+
+## Contents
+
+- `iscene_config.json`: public release metadata and checkpoint layout.
+- `config.yml`: inference-only architecture configuration.
+- `ckpts/iscene_denoiser.pt`: IScene denoiser checkpoint.
+- `ckpts/iscene_image_conditioner.pt`: IScene image-conditioner checkpoint.
+
+## Usage
+
+The public IScene code should load this package with:
+
+```python
+from iscene.inference.inferencer import ISceneInferencer
+
+inferencer = ISceneInferencer.from_pretrained("LuLing/IScene")
+```
+
+For local testing before uploading to Hugging Face:
+
+```python
+inferencer = ISceneInferencer.from_pretrained("release_hf/IScene-v1")
+```
+
+## Notes
+
+- This package contains only the IScene release checkpoint files, not historical training logs or experimental checkpoints.
+- The IScene code is expected to load the TRELLIS base model from `microsoft/TRELLIS-image-large` or an equivalent local mirror.
+
+## Attribution
+
+IScene-v1 builds on the TRELLIS image-conditioned 3D generation backbone. The public loader uses TRELLIS base components from `microsoft/TRELLIS-image-large`.
+
+Please also cite and respect the license terms of TRELLIS:
+
+- Project: https://trellis3d.github.io/
+- Code: https://github.com/microsoft/TRELLIS
+- Model: https://huggingface.co/microsoft/TRELLIS-image-large
+- Paper: Structured 3D Latents for Scalable and Versatile 3D Generation
+
+## License
+
+This model package is prepared for release under the MIT License. Third-party TRELLIS attribution is included above and should be preserved.

ckpts/.gitkeep

@@ -0,0 +1 @@
+

ckpts/iscene_denoiser.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e4642c719615ac035e23b976220b26af22992d3744c361624a17dbdc10b96c9
+size 2239036066

ckpts/iscene_image_conditioner.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a5d6d5d527edd6c5e95ecbc28e21ffb9bfdec08a725eaf52e5df8141c7a35dd
+size 1217639882

config.yml

@@ -0,0 +1,34 @@
+# Inference-only architecture config for IScene-v1.
+# This file intentionally excludes training logs, data paths, cluster settings,
+# experiment names, and checkpoint metadata.
+
+models:
+  denoiser:
+    name: SparseStructureSceneContextFlowModel
+    args:
+      resolution: 16
+      in_channels: 8
+      out_channels: 8
+      model_channels: 1024
+      cond_channels: 1024
+      num_blocks: 24
+      num_heads: 16
+      mlp_ratio: 4
+      patch_size: 1
+      pe_mode: ape
+      qk_rms_norm: true
+      use_fp16: true
+      scene_context_attn_num: 5
+      learning_pattern: full-finetune
+      exp_setting: global
+
+  img_conditioner:
+    name: ImageConditioner
+    args:
+      image_cond_model: dinov2_vitl14_reg
+      cond_in_channels: 3
+      use_fp16: false
+
+dataset:
+  args:
+    exp_setting: global

iscene_config.json

@@ -0,0 +1,7 @@
+{
+  "model_name": "IScene-v1",
+  "base_model_id": "microsoft/TRELLIS-image-large",
+  "config_file": "config.yml",
+  "denoiser_checkpoint": "ckpts/iscene_denoiser.pt",
+  "image_conditioner_checkpoint": "ckpts/iscene_image_conditioner.pt"
+}