How aiuniai/unique3d works
Unique3D turns a single input image into a textured 3D mesh. The main Gradio workflow optionally removes the background and upscales small images, predicts multiple views, reconstructs and refines geometry, then exports a mesh and optional preview video. A separate stepwise UI exposes multiview generation and mesh reconstruction, including optional supplied normals; the README also documents local and hosted demos and ComfyUI support. The graph follows explicit calls in sampled UI and prediction code; reconstruction internals and several documented integrations are represented only at their supported directory paths.
Main entry points: gradio_3dgen.py Full-run Gradio · gradio_3dgen_steps.py Stepwise Gradio · mvimg_prediction.py Multiview prediction · multiview_inference.py Geometry reconstruction
Components 19
User interfaces
- Full-run Gradio
app/gradio_3dgen.py - Stepwise Gradio
app/gradio_3dgen_steps.py
Image generation
- Background removal
app/custom_models/mvimg_prediction.py - Image preprocessing
scripts/utils.py - Image upscaling
scripts/refine_lr_to_sr.py - Multiview prediction
app/custom_models/mvimg_prediction.py - Normal prediction
app/custom_models/normal_prediction.py
Mesh reconstruction
- Geometry reconstruction
scripts/multiview_inference.py - Mesh reconstruction
mesh_reconstruction - Adaptive remeshing
mesh_reconstruction/opt.py - Mesh and video export
scripts/utils.py
Diffusion internals
- Diffusion training
custum_3d_diffusion/trainings - Image diffusion pipelines
custum_3d_diffusion/custum_pipeline - Attention processors
custum_3d_diffusion/custum_modules/attention_processors.py - Configurable UNet
custum_3d_diffusion/custum_modules/unifield_processor.py
Other components
- Image user
- Hosted demos
- ComfyUI integration
- Generated mesh
Connections 21
- Image user → Full-run Gradio: submits image
- Image user → Stepwise Gradio: submits image
- Full-run Gradio → Image upscaling: upscales small image
- Full-run Gradio → Multiview prediction: predicts views
- Multiview prediction → Background removal: removes background
- Multiview prediction → Image preprocessing: preprocesses image
- Multiview prediction → Diffusion training: runs pipeline forward
- Diffusion training → Image diffusion pipelines: constructs pipeline
- Image diffusion pipelines → Configurable UNet: uses UNet
- Configurable UNet → Attention processors: installs processors
- Full-run Gradio → Geometry reconstruction: reconstructs geometry
- Geometry reconstruction → Mesh reconstruction: builds mesh
- Mesh reconstruction → Adaptive remeshing: optimizes topology
- Full-run Gradio → Mesh and video export: exports mesh
- Mesh and video export → Generated mesh: writes result
- Stepwise Gradio → Multiview prediction: generates views
- Stepwise Gradio → Normal prediction: optional normal prediction
- Stepwise Gradio → Geometry reconstruction: reconstructs mesh
- Stepwise Gradio → Mesh and video export: exports mesh
- Hosted demos → Image user: offers inference
- ComfyUI integration → Image user: offers workflow