DICEPTION¶

Development line: project:diception · thread diception-development
Last event: 2025-10-02 · 2 dated since 2025-04-21 · Researched: 2026-09-04 · confidence: medium

What it is¶

DICEPTION is an SD3-based diffusion model for practitioners who want one local model for visual-perception outputs.

• Monocular depth and surface normals.
• Pose estimation.
• Entity, semantic, and point-prompted segmentation.

Development line¶

• 2025-04-21 — DICEPTION project resources were publicly linked. On 2025-04-21, links appeared for the project website, its GitHub repository, and a Hugging Face Spaces demo. They established a dated public entry point for the project, its source code, and an interactive demonstration.
• 2025-10-02 — DICEPTION Hugging Face repository was added. On 2025-10-02, links added the DICEPTION Hugging Face repository alongside the earlier project post. This added a distinct model-repository access point to the dated public trail, without showing which files changed or whether this was a first release.

What changed¶

DICEPTION development:

• 2025-02-24/25 — arXiv v1/v2 established the diffusion-based generalist-perception method.
• 2025-04-21 — public entry points were the project page, repository, and web demo. Usable inference code and weights came later in September, so local install did not work yet.
• 2025-09-19 — accepted as a NeurIPS 2025 Spotlight.
• 2025-09-21 — maintainers released inference code and the v1 model.
• 2025-10-02 — files added the Hugging Face model tree with the DICEPTION_v1.pth checkpoint.
• 2026-09-04 — the hosted demo reports RUNTIME_ERROR, so local code plus weights is the usable route.

How to use this¶

From 2025-10-02, consult the DICEPTION project site, repository, demo, and Hugging Face repository together when evaluating the project.

1. Clone the repository, create its Python 3.10 Conda environment, and install requirements. — https://github.com/aim-uofa/Diception
2. Download Stable Diffusion 3 Medium in Diffusers format and DICEPTION_v1.pth, then set --pretrained_model_path and --diception_path to those local directories. — https://huggingface.co/Canyu/DICEPTION/tree/main
3. Run inference.py with an input image and an exact task token such as [[image2depth]], [[image2normal]], [[image2pose]], [[image2entity]], or [[image2segmentation]]. — https://github.com/aim-uofa/Diception/blob/main/inference.py
4. For semantic segmentation, append the COCO category to [[image2semantic]]; for interactive segmentation, pass normalized y,x points, with at most five points. — https://github.com/aim-uofa/Diception
5. For datasets, use batch_inference.py with the supplied JSON shape; use --save_npy when depth or normal values are needed rather than only a visualization. — https://github.com/aim-uofa/Diception

Best practices¶

• Provision a CUDA-capable bfloat16 environment before running the released script: it explicitly moves the model to CUDA/bfloat16 and loads the checkpoint onto CUDA. — https://github.com/aim-uofa/Diception/blob/main/inference.py
• Start with the maintainers' 28 inference steps and guidance scale 2.0 for general quality; use 1 step and guidance 1.0 only when speed matters, especially for depth or normals. — https://github.com/aim-uofa/Diception
• Validate at the task's native evaluation format: single-image inference always works at 768×768 and saves a side-by-side input/output image, not an untouched native-resolution prediction. — https://github.com/aim-uofa/Diception/blob/main/inference.py
• Download the v1 .pth only from the named model repository and treat it as a PyTorch pickle artifact; the Hub flags pickle imports and the code uses torch.load. — https://huggingface.co/Canyu/DICEPTION/tree/main
• Do not make the hosted Space a dependency of a workflow until it is repaired; use the released local code and weights instead. — https://huggingface.co/spaces/Canyu/Diception-Demo
• For commercial use, follow the project's instruction to contact the stated maintainer rather than assuming the academic-use note covers deployment. — https://github.com/aim-uofa/Diception

Superseded by this¶

• 2025-04-21 research/demo-only access: superseded on 2025-09-21 by released local inference code and v1 weights.
• 2025-04-21 demo-first access: obsolete as a current workflow on 2026-09-04 because the hosted Hugging Face Space reports RUNTIME_ERROR.

Still unknown¶

• We have no original post text for the 2025-04-21 and 2025-10-02 items, so their wording can only be reconstructed from dated links.
• We found secondary translated summaries but no first-party Chinese operating documentation or independent Chinese execution report; none was used as practical evidence.
• We found no current local end-to-end run, VRAM measurement, or supported-hardware matrix. CUDA/bfloat16 code is implementation evidence, not a passing runtime receipt.
• The current README still lists training and few-shot fine-tuning code as planned, so the paper's 50-image and 1%-parameter adaptation result is not a runnable public workflow.

Sources¶

Agent brief¶

• Subject: project:diception, thread diception-development, 2 dated events 2025-04-21 → 2025-10-02.
• Practical note: From 2025-10-02, practitioners should consult the DICEPTION project site, source repository, demo, and linked Hugging Face repository together when evaluating the project or locating its public artifacts.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.