teamwork-release

<h1 align="center">Teamwork - Models Release</h1> <p align="center"> <a href="https://bin.samsartor.com/teamwork.pdf">📃 Paper</a> | <a href="https://samsartor.com/teamwork">🌐 Website</a> | <a href="https://bin.samsartor.com/teamwork_supplemental.pdf">🖼️ Supplemental</a> </p> <p align="center"><img width="700px" src="https://samsartor.com/teamwork_teaser.svg" /></p>

Models presented in the paper "Teamwork: Collaborative Diffusion with Low-rank Coordination and Adaptation".

Use with https://github.com/samsartor/teamwork.

from teamwork.pipelines import TeamworkPipeline
from PIL import Image

pipe = TeamworkPipeline.from_checkpoint(
    'samsartor/teamwork-release',
    'decomposition_heterogeneous_sd3.safetensors',
).to('cuda')

generated = pipe({'image': Image.open('./demo/red_glass_sphere.png'})

Teamwork's safetensors files contain (roughly) the parameters of LoRAs to apply to a base model. However, in addition to adapting the base model for a new task, Teamwork enables the model to accept any number of input images and then collaboratively generate any number of outputs. The available models are:

<table> <th><td>Description</td><td>Base Model</td><td>Resolution</td><td>Inputs</td><td>Outputs</td></th> <tr> <td>decomposition_heterogeneous_sd3.safetensors</td> <td>Intrinsic image decomposition trained on heterogeneous datasets<br>(InteriorVerse, HyperSim, CGIntrinsics, Infinigen, ABC).</td> <td>SD3</td> <td>1024x1024</td> <td>image</td> <td>diffuse<br>specular<br>roughness<br>normals<br>depth<br>albedo<br>inverseshading<br>diffuseshading<br>residual</td> </tr> <tr> <td>decomposition_interiorverse_sd3.safetensors</td> <td>Intrinsic image decomposition specialized for InteriorVerse dataset.</td> <td>SD3</td> <td>640x480</td> <td>image</td> <td>diffuse<br>specular<br>roughness<br>normals<br>depth<br>albedo<br>inverseshading</td> </tr> <tr> <td>inpainting_sd3.safetensors</td> <td>Text-conditional inpainting of masked regions.</td> <td>SD3</td> <td>1024x1024</td> <td>mask<br>masked</td> <td>image</td> </tr> <tr> <td>neuralshading_sd3.safetensors</td> <td>Synthesizes images from intrinsic components under unknown random lighting.</td> <td>SD3</td> <td>1024x1024</td> <td>albedo<br>diffuse<br>specular<br>roughness<br>normals<br>depth</td> <td>image<br>diffuseshading<br>inverseshading<br>residual</td> </tr> <tr> <td>colocated_svbrdf_estimation_flux_256.safetensors</td> <td>Estimates material properties from a single flash-lit photograph.</td> <td>FLUX.1-dev</td> <td>256x256</td> <td>image<br>halfway</td> <td>diffuse<br>specular<br>roughness<br>normals</td> </tr> <tr> <td>colocated_svbrdf_estimation_flux_512.safetensors</td> <td>Estimates material properties from a single flash-lit photograph.</td> <td>FLUX.1-dev</td> <td>512x512</td> <td>image<br>halfway</td> <td>diffuse<br>specular<br>roughness<br>normals</td> </tr> <tr> <td>colocated_svbrdf_estimation_sd3_256.safetensors</td> <td>Estimates material properties from a single flash-lit photograph.</td> <td>SD3</td> <td>256x256</td> <td>image<br>halfway</td> <td>diffuse<br>specular<br>roughness<br>normals</td> </tr> <tr> <td>colocated_svbrdf_estimation_sd3_512.safetensors</td> <td>Estimates material properties from a single flash-lit photograph.</td> <td>SD3</td> <td>512x512</td> <td>image<br>halfway</td> <td>diffuse<br>specular<br>roughness<br>normals</td> </tr> <tr> <td>colocated_svbrdf_estimation_sdxl_256.safetensors</td> <td>Estimates material properties from a single flash-lit photograph.</td> <td>SDXL</td> <td>256x256</td> <td>image<br>halfway</td> <td>diffuse<br>specular<br>roughness<br>normals</td> </tr> </table>

Citation

@conference{Sartor:2025:TCD,
    author    = {Sartor, Sam and Peers, Pieter},
    title     = {Teamwork: Collaborative Diffusion with Low-rank Coordination and Adaptation},
    month     = {December},
    year      = {2025},
    booktitle = {ACM SIGGRAPH Asia Conference Proceedings},
}