AI Models Referenced in This Study¶
| Model | Category | Description |
|---|---|---|
| AAGNet | Machine Learning (ML) | Machining-feature recognition on B-rep adjacency graphs. |
| AB-UPT | Machine Learning (ML) | Open transformer architecture for physics surrogates (Emmi AI). |
| Aleph Alpha | Language model (L) | German sovereign-AI LLM provider (Pharia models). |
| BGE-M3 | Knowledge (K) | Multilingual text-embedding model for retrieval. |
| BRepNet | Machine Learning (ML) | Face segmentation / feature recognition on B-rep topology. |
| Claude | Language model (L) | Anthropic's proprietary general-purpose LLM family. |
| Deep Ritz | Machine Learning (ML) | Neural PDE solver via energy minimisation (method). |
| DeepONet | Machine Learning (ML) | Neural operator learning function-to-function maps (method). |
| DeepSeek | Language model (L) | Open-weight LLM family, including reasoning models. |
| DINOv2 | Knowledge (K) | Self-supervised image feature / embedding model. |
| FNO | Machine Learning (ML) | Neural operator solving PDEs in Fourier space (method). |
| Gemma | Language model (L) | Google's lightweight open-weight LLM family. |
| Google Gemini | Language model (L) | Google's proprietary multimodal LLM family. |
| InternVL3 | Knowledge (K) | Open vision-language model (image + text → text). |
| jina-v3 | Knowledge (K) | Multilingual text-embedding model with task adapters (non-commercial). |
| Llama | Language model (L) | Meta's open-weight general-purpose LLM family. |
| MeshGraphNets | Machine Learning (ML) | GNN that learns mesh-based physical simulation (method). |
| Microsoft / GitHub Copilot | Language model (L) | Coding-and-text assistant wrapping OpenAI LLMs. |
| Mistral | Language model (L) | Open-and-commercial family of general-purpose text LLMs. |
| Mistral "Large Industry Model" (LIM) | Machine Learning (ML) | Proprietary crash-simulation surrogate (BMW × Mistral). |
| Mistral Small 3 | Language model (L) | Compact open-weight general-purpose LLM. |
| Molmo | Knowledge (K) | Open vision-language model with region pointing. |
| Nemotron | Language model (L) | NVIDIA's open LLM family. |
| NeuralDEM | Machine Learning (ML) | Neural surrogate for discrete-element (particle) simulation (Emmi AI). |
| NeuralMould | Machine Learning (ML) | Proprietary injection-moulding simulation surrogate. |
| Nomic-embed | Knowledge (K) | Open, reproducible text-embedding model. |
| NV-Embed-v2 | Knowledge (K) | LLM-based text-embedding model (non-commercial). |
| NVIDIA Cosmos | Machine Learning (ML) | World model generating physics-plausible video for training. |
| OpenAI GPT / ChatGPT | Language model (L) | OpenAI's proprietary general-purpose LLM and chat product. |
| OpenCLIP | Knowledge (K) | Open image–text shared-embedding model (CLIP reproduction). |
| OpenShape | Knowledge (K) | Open text–image–point-cloud 3D embedding model. |
| Phi-4 | Language model (L) | Microsoft's small open-weight reasoning LLM. |
| PINN | Machine Learning (ML) | Neural PDE solver via physics-in-the-loss (method). |
| Project Bernini | Machine Learning (ML) | Experimental generative-3D shape model (research only). |
| Qwen2.5-VL | Knowledge (K) | Alibaba's open vision-language model. |
| Qwen3 / Qwen | Language model (L) | Alibaba's open-weight general-purpose LLM family. |
| Qwen3-Embedding | Knowledge (K) | Alibaba's text-embedding model for retrieval. |
| Stable Diffusion | Machine Learning (ML) | Open text-to-image latent diffusion model. |
| ULIP-2 | Knowledge (K) | Image–language–point-cloud 3D pretraining (point clouds). |
| Uni3D | Knowledge (K) | Large-scale open 3D point-cloud embedding model. |
| UV-Net | Knowledge / Machine Learning (K/ML) | GNN that embeds B-rep CAD solids for retrieval and segmentation. |
| UVStyle-Net | Knowledge / Machine Learning (K/ML) | Style-similarity search over B-rep CAD solids. |
Mistral¶
Creator: Mistral AI (Paris, France; founded 2023)
Input: natural-language text (prompt tokens); some variants also accept images.
Output: natural-language text (generated tokens).
Architecture (code) available? Partly
— the open-weight models (Mistral 7B, Mixtral 8×7B) ship with an Apache-2.0 licence and inference code; the flagship Mistral Large is closed and API-only. The underlying transformer design is standard and published.
Trained weights available? Partly
— 7B / Mixtral / Small weights are downloadable under Apache-2.0; Mistral Large weights are not released.
Purpose (as created): a family of autoregressive large language models that generate and reason over natural-language text for chat, coding and general text tasks.
Mistral Small 3¶
Creator: Mistral AI (Paris, France).
Input: natural-language text.
Output: natural-language text.
Architecture (code) available? Yes
— open weights with a standard transformer design, released under Apache-2.0.
Trained weights available? Yes
— downloadable under Apache-2.0.
Purpose (as created): a compact general-purpose language model optimised for low-latency text generation and reasoning.
OpenAI GPT / ChatGPT¶
Creator: OpenAI (San Francisco, USA).
Input: natural-language text; later models (GPT-4o onward) also accept images and audio.
Output: natural-language text (and, in multimodal variants, images/audio).
Architecture (code) available? No
— proprietary; neither the architecture details nor the code are released.
Trained weights available? No
— closed; usable only through OpenAI's API and products.
Purpose (as created): a family of autoregressive general-purpose language models for text generation, reasoning and task completion; ChatGPT is the conversational product built on them.
Microsoft / GitHub Copilot¶
Creator: Microsoft / GitHub (USA; underlying models by OpenAI).
Input: natural-language text plus working context — for GitHub Copilot, the surrounding source code and comments; for Microsoft Copilot, documents and app context.
Output: natural-language text and code completions/suggestions.
Architecture (code) available? No
— proprietary product.
Trained weights available? No
— closed; the models are served, not distributed.
Purpose (as created): an AI assistant embedded inside Microsoft and GitHub tools that generates code and text suggestions from natural-language and code context.
Claude¶
Creator: Anthropic (San Francisco, USA).
Input: natural-language text; images from Claude 3 onward; large context windows.
Output: natural-language text.
Architecture (code) available? No
— proprietary.
Trained weights available? No
— closed; API and product only.
Purpose (as created): a family of general-purpose language models for conversation, reasoning, coding and document analysis, trained with an emphasis on helpfulness and safety.
Google Gemini¶
Creator: Google DeepMind (USA / UK).
Input: multimodal — text, images, audio, video and code.
Output: natural-language text (and structured/multimodal output in some variants).
Architecture (code) available? No
— proprietary; its openly-released sibling is Gemma.
Trained weights available? No
— closed; API and product only.
Purpose (as created): a family of natively multimodal general-purpose language models for reasoning, generation and tool use.
Llama¶
Creator: Meta AI (USA).
Input: natural-language text; vision variants (Llama 3.2+) also accept images.
Output: natural-language text.
Architecture (code) available? Yes
— the model architecture and reference inference code are published.
Trained weights available? Yes, with restrictions
— downloadable under Meta's custom Llama Community License (not OSI-approved; carries usage and redistribution conditions).
Purpose (as created): a family of open-weight general-purpose language models released by Meta for research and commercial use.
DeepSeek¶
Creator: DeepSeek AI (Hangzhou, China).
Input: natural-language text and code.
Output: natural-language text; reasoning variants emit an explicit chain of thought followed by the answer.
Architecture (code) available? Yes
— technical reports and code released.
Trained weights available? Yes
— downloadable (MIT licence on most releases).
Purpose (as created): a family of open-weight language models, including reasoning-specialised ones, for text generation, coding and mathematical reasoning.
Phi-4¶
Creator: Microsoft Research (USA).
Input: natural-language text.
Output: natural-language text.
Architecture (code) available? Yes
— open weights with a standard transformer design, MIT-licensed.
Trained weights available? Yes
— downloadable under MIT.
Purpose (as created): a small (~14B-parameter) language model trained heavily on curated and synthetic data to reach strong reasoning at compact size.
Qwen3 / Qwen¶
Creator: Alibaba Cloud (Qwen team; China).
Input: natural-language text (vision-language variants add images).
Output: natural-language text.
Architecture (code) available? Yes
— technical reports and code released.
Trained weights available? Yes
— most sizes under Apache-2.0; the largest (72B) under a custom Qwen licence.
Purpose (as created): a family of open-weight general-purpose language models for text generation, reasoning, coding and multilingual tasks.
Gemma¶
Creator: Google DeepMind (USA / UK).
Input: natural-language text (multimodal variants also accept images).
Output: natural-language text.
Architecture (code) available? Yes
— open weights with reported architecture.
Trained weights available? Yes
— downloadable under the custom Gemma Terms of Use.
Purpose (as created): a family of lightweight open-weight language models built from the same research as Gemini.
Nemotron¶
Creator: NVIDIA (USA).
Input: natural-language text.
Output: natural-language text.
Architecture (code) available? Yes
— released with technical reports.
Trained weights available? Yes
— downloadable under the NVIDIA Open Model License.
Purpose (as created): a family of open language models tuned for reasoning, chat and synthetic-data generation.
Aleph Alpha¶
Creator: Aleph Alpha (Heidelberg, Germany; founded 2019).
Input: natural-language text (multilingual); some variants are multimodal.
Output: natural-language text.
Architecture (code) available? Partly
— the Pharia-1 models were released with open weights and model code; the wider PhariaAI enterprise stack is proprietary.
Trained weights available? Partly
— Pharia-1-LLM weights are downloadable under Aleph Alpha's own Open Aleph License (research / non-commercial); production enterprise deployments are separately licensed.
Purpose (as created): a family of enterprise LLMs built for on-premises, data-sovereign deployment in regulated European industry. Cited in the study as the one on-premises (non-rented) model in a Tier-1's stack (Bosch).
DINOv2¶
Creator: Meta AI Research (USA).
Input: images (RGB).
Output: image embedding vectors (global and dense/patch-level feature representations).
Architecture (code) available? Yes
— open-source code at github.com/facebookresearch/dinov2.
Trained weights available? Yes
— pretrained weights downloadable under Apache-2.0.
Purpose (as created): a self-supervised vision model that learns general-purpose image feature representations from unlabelled images, usable without fine-tuning for classification, retrieval, segmentation and depth estimation.
OpenCLIP¶
Creator: open-source community (LAION, based in Germany, and contributors); a reproduction of OpenAI's CLIP.
Input: images and/or natural-language text.
Output: image and text embeddings placed in a single shared vector space.
Architecture (code) available? Yes
— open-source (MIT) at github.com/mlfoundations/open_clip.
Trained weights available? Yes
— many pretrained checkpoints downloadable (trained largely on LAION datasets).
Purpose (as created): an open reproduction of CLIP (Contrastive Language–Image Pre-training) that maps images and text into a shared embedding space, enabling zero-shot image classification and image–text (cross-modal) search.
InternVL3¶
Creator: OpenGVLab (Shanghai AI Laboratory, China) with academic and industry partners.
Input: images (and video) plus natural-language text.
Output: natural-language text (answers, descriptions); also usable to produce multimodal embeddings.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes
— downloadable under MIT.
Purpose (as created): an open vision-language model that reads images and text and generates textual responses — for visual question answering, image description and document understanding.
Molmo¶
Creator: Allen Institute for AI (Ai2; Seattle, USA).
Input: images plus natural-language text.
Output: natural-language text, with the ability to "point" to image regions.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes
— downloadable under Apache-2.0.
Purpose (as created): an open vision-language model that answers questions about images and can indicate specific image regions, built on openly-released training data.
Qwen2.5-VL¶
Creator: Alibaba Cloud (Qwen team; China).
Input: images (and video) plus natural-language text.
Output: natural-language text, including bounding boxes and structured document extraction.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes
— smaller sizes under Apache-2.0; the 72B under a custom Qwen licence.
Purpose (as created): the vision-language member of the Qwen family — reads images, documents and video with a text prompt and produces textual answers, localisation and structured extraction.
BGE-M3¶
Creator: Beijing Academy of Artificial Intelligence (BAAI; China).
Input: natural-language text (multilingual; supports long documents).
Output: text embedding vectors; supports dense, sparse (lexical) and multi-vector retrieval at once.
Architecture (code) available? Yes
— open-source (part of BAAI's FlagEmbedding).
Trained weights available? Yes
— downloadable under MIT.
Purpose (as created): a text-embedding model that converts text into vectors for semantic search and retrieval; "M3" denotes its multi-function, multi-lingual and multi-granularity design.
Nomic-embed¶
Creator: Nomic AI (USA).
Input: natural-language text (a vision variant handles images).
Output: text embedding vectors.
Architecture (code) available? Yes
— open-source, with the training data and recipe also released.
Trained weights available? Yes
— downloadable under Apache-2.0.
Purpose (as created): a fully-reproducible open text-embedding model that maps text into vectors for semantic search and retrieval.
Qwen3-Embedding¶
Creator: Alibaba Cloud (Qwen team; China).
Input: natural-language text (multilingual).
Output: text embedding vectors.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes
— downloadable under Apache-2.0.
Purpose (as created): the embedding member of the Qwen3 family — converts text into vectors for semantic search, retrieval and reranking.
NV-Embed-v2¶
Creator: NVIDIA (USA).
Input: natural-language text.
Output: text embedding vectors.
Architecture (code) available? Yes
— research code and model card published.
Trained weights available? Yes, non-commercial only
— downloadable under CC-BY-NC-4.0, which prohibits commercial use.
Purpose (as created): a text-embedding model built on a large-language-model backbone that produces vectors for retrieval tasks; released for research.
jina-v3¶
Creator: Jina AI (Berlin, Germany).
Input: natural-language text (multilingual, long context).
Output: text embedding vectors, specialised per task through swappable adapters.
Architecture (code) available? Yes
— model code and card published.
Trained weights available? Yes, non-commercial only
— weights under CC-BY-NC-4.0 (commercial use is via Jina's API/licence).
Purpose (as created): a multilingual text-embedding model with task-specific adapters (for queries, documents and classification) that produces vectors for semantic search and retrieval.
OpenShape¶
Creator: academic collaboration (Liu et al.; UC San Diego and partners), 2023.
Input: 3D point clouds1 (points sampled from a mesh), aligned with images and text.
Output: 3D shape embedding vectors in a shared text–image–point-cloud space.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes, but use-restricted
— pretrained checkpoints are published on Hugging Face under OpenRAIL (a responsible-AI licence that permits use but attaches behavioural restrictions); downloadable for experiments, not a clean commercial licence.
Purpose (as created): scaling multimodal (text–image–point-cloud) representation learning so 3D shapes can be classified and retrieved open-vocabulary and zero-shot.
ULIP-2¶
Creator: Salesforce Research (USA), 2023.
Input: 3D point clouds1 with paired image and language.
Output: point-cloud embeddings aligned into an image–text space.
Architecture (code) available? Yes
— open-source, code under BSD-3-Clause.
Trained weights available? Yes, licence unclear
— checkpoints published on Hugging Face and downloadable, but the BSD-3-Clause covers the code only; no licence is stated for the weights themselves, and the training data (Objaverse under ODC-BY, plus ShapeNet terms) adds its own conditions.
Purpose (as created): unified image–language–point-cloud pretraining that produces 3D representations for classification and retrieval, using model-generated descriptions to scale training.
Uni3D¶
Creator: Beijing Academy of Artificial Intelligence (BAAI; China), 2023.
Input: 3D point clouds1 (with image / text alignment).
Output: large-scale 3D point-cloud embeddings in a shared space.
Architecture (code) available? Yes
— open-source, code under MIT.
Trained weights available? Yes, licence unclear
— checkpoints (Uni3d-B/L/g) published on Hugging Face and downloadable, but the MIT licence covers the code only; no licence is stated for the weights themselves, and they are trained on ensembled Objaverse data that carries its own terms.
Purpose (as created): scaling a unified 3D representation to large model and data sizes for open-world 3D understanding, classification and retrieval.
UV-Net¶
Creator: Autodesk AI Lab (USA; Jayaraman et al., CVPR 2021).
Input: a B-rep solid (read from STEP): each face is sampled into a 2D UV-grid of surface points and normals, and the faces are linked into a face-adjacency graph.
Output: learned embedding vectors — per-face and per-solid — used downstream for classification, retrieval and segmentation.
Architecture (code) available? Yes
— open-source, MIT-licensed, at github.com/AutodeskAILab/UV-Net.
Trained weights available? No (no general-purpose model)
— only demo checkpoints on narrow datasets (e.g. fonts, machining features) are published; nothing moulding-ready.
Purpose (as created): a graph neural network that learns vector representations of boundary-representation (B-rep) CAD models — combining surface geometry (UV-grids) with topology (the face graph) — for shape classification, retrieval and segmentation.
UVStyle-Net¶
Creator: Autodesk AI Lab (USA; Meltzer et al., ICCV 2021).
Input: a B-rep solid, encoded with a UV-Net-style representation.
Output: style-similarity scores/embeddings (derived from feature Gram matrices) between CAD solids.
Architecture (code) available? Yes, but non-commercial
— open-source, but released under CC BY-NC-SA 4.0 (non-commercial), unlike UV-Net's MIT. This blocks commercial reuse of the released code; the method can be reimplemented on the MIT-licensed UV-Net encoder.
Trained weights available? Partly
— code and reference models are published (non-commercial); no general production-ready weights.
Purpose (as created): a method, built on UV-Net, that measures stylistic similarity between B-rep CAD solids so a library can be searched by visual/geometric style rather than by function.
BRepNet¶
Creator: Autodesk AI Lab (USA; Lambourne et al., CVPR 2021).
Input: a B-rep solid, processed through its topology — "coedge" walks over faces, edges and coedges, plus per-entity geometric features.
Output: per-face segmentation labels (e.g. machining-feature classes).
Architecture (code) available? Yes
— open-source, released with the Fusion 360 Gallery dataset.
Trained weights available? Partly
— code and trained models for the Fusion 360 Gallery segmentation task are published.
Purpose (as created): a neural network that performs message passing directly on B-rep topology (coedges) to segment and classify the faces of a CAD solid — i.e. to recognise machined features.
AAGNet¶
Creator: academic (Wu et al., China; 2023); released open-source.
Input: a B-rep solid represented as an attributed adjacency graph — faces as nodes carrying geometric attributes, edges as their adjacency.
Output: machining-feature recognition — per-face labels plus instance segmentation grouping faces into features.
Architecture (code) available? Yes
— open-source, MIT-licensed.
Trained weights available? Partly
— code and models trained on machining-feature datasets are published.
Purpose (as created): a graph neural network that recognises manufacturing (machining) features in a B-rep CAD model by learning on its attributed adjacency graph, simultaneously classifying faces and grouping them into feature instances.
PINN — Physics-Informed Neural Network¶
Creator: Raissi, Perdikaris & Karniadakis (Brown University, USA), 2019.
Input: spatial/temporal coordinates (plus any PDE parameters and boundary/initial conditions); training uses the governing equation's residual rather than labelled solution data.
Output: the predicted physical field value(s) at the queried coordinates — a continuous approximation of the PDE's solution.
Architecture (code) available? Yes
— a published method with many open implementations (not a single owned model).
Trained weights available? Not applicable
— trained per problem; no distributed pretrained model.
Purpose (as created): a neural network trained to satisfy a differential equation by embedding the PDE and its boundary/initial conditions into the loss function, so the network itself becomes a mesh-free solver of that equation.
DeepONet¶
Creator: Lu, Jin & Karniadakis (Brown University, USA), 2021.
Input: two parts — a sampled input function (e.g. a loading or boundary condition, fed to a "branch" network) and a query location (fed to a "trunk" network).
Output: the output function's value at the query location — i.e. it learns an operator mapping functions to functions.
Architecture (code) available? Yes
— a published method with open implementations.
Trained weights available? Not applicable
— trained per operator / problem family.
Purpose (as created): a neural-network architecture that learns operators — mappings from one function to another (e.g. from an input condition to a PDE's solution field) — so one trained network can address a whole family of problems, not a single instance.
FNO — Fourier Neural Operator¶
Creator: Li et al. (Caltech, USA), 2020.
Input: a discretised input field on a grid (e.g. an initial condition or coefficient field).
Output: the predicted output field on the grid (the PDE's solution), largely independent of grid resolution.
Architecture (code) available? Yes
— a published method with open code.
Trained weights available? Not applicable
— trained per problem family.
Purpose (as created): a neural operator that learns mappings between function spaces by performing its convolutions in the Fourier domain, letting one trained model approximate a PDE's solution operator across resolutions.
Deep Ritz¶
Creator: Weinan E (Princeton University, USA) & Bing Yu (Peking University, China), 2018.
Input: spatial coordinates (sample points across the domain).
Output: the predicted solution field, obtained by minimising the PDE's variational (energy) functional.
Architecture (code) available? Yes
— a published method.
Trained weights available? Not applicable
— trained per problem.
Purpose (as created): a neural-network method that solves partial differential equations by recasting them as an energy-minimisation (Ritz variational) problem, with the network representing the trial solution.
MeshGraphNets¶
Creator: DeepMind (Pfaff, Fortunato et al.), 2021.
Input: a simulation mesh — nodes and edges with per-node state — i.e. the domain's mesh graph.
Output: the predicted next-step physical state at every mesh node (e.g. deflection, flow), rolled forward over time.
Architecture (code) available? Yes
— a published method with open reference code (not a single owned model).
Trained weights available? Not applicable
— trained per problem family; no distributed general model.
Purpose (as created): a graph-neural-network method that learns mesh-based physical simulation, predicting how each node of a mesh evolves. Noted in the study as a natural fit for predicting deflection at every node of a part.
AB-UPT¶
Creator: Emmi AI (Linz, Austria), 2025; released open-source.
Input: the geometry and physical setup of a simulation (mesh / point sampling of the domain).
Output: predicted physical fields — a fast surrogate for a numerical solver.
Architecture (code) available? Yes
— the Anchored-Branched Universal Physics Transformer architecture was open-sourced on GitHub.
Trained weights available? No
— the trained domain model (NeuralMould) is kept closed; only the architecture is open.
Purpose (as created): a transformer architecture for large-scale physics-simulation surrogates; the open half of the SIMCON / Emmi AI moulding-surrogate story.
NeuralDEM¶
Creator: Emmi AI / JKU Linz (Austria), 2024–2025.
Input: the state of a discrete-element (particle) simulation.
Output: the predicted evolution of the particle system — a fast surrogate for DEM simulation.
Architecture (code) available? Yes
— published, and open-sourced by Emmi AI alongside the AB-UPT architecture.
Trained weights available? No
— trained models are not distributed.
Purpose (as created): a neural surrogate for the Discrete Element Method (large-scale particle physics), demonstrating near-real-time replacement of a slow numerical simulation.
Mistral "Large Industry Model" (LIM)¶
Creator: Mistral AI (France) with BMW Group (Germany), 2026.
Input: (vendor-disclosed, internals not public) large-scale crash-simulation data — simulation setups and results.
Output: (inferred) fast surrogate predictions of crash-simulation outcomes.
Architecture (code) available? No
— proprietary; design not disclosed.
Trained weights available? No
— private to BMW and Mistral.
Purpose (as created): a large domain-specific model trained on BMW's crash-simulation corpus to act as a fast learned surrogate for physical (crash) simulation.
NeuralMould¶
Creator: Emmi AI (Linz, Austria; later acquired by Mistral) with SIMCON (Germany), 2026.
Input: (inferred) an injection-moulding case — part geometry and process settings — using SIMCON's solver and training corpus.
Output: (inferred) predicted moulding results (e.g. filling behaviour), as a fast surrogate for the numerical solver.
Architecture (code) available? No
— proprietary.
Trained weights available? No
— proprietary.
Purpose (as created): a domain-specific "Large Engineering Model" trained to act as a learned surrogate for injection-moulding simulation (mould filling), accelerating or replacing the numerical solver.
NVIDIA Cosmos¶
Creator: NVIDIA (USA), 2025.
Input: text prompts and/or images/video used as conditioning.
Output: generated video predicting a scene's future states (synthetic sensor data).
Architecture (code) available? Yes
— released with code.
Trained weights available? Yes
— downloadable under the NVIDIA Open Model License.
Purpose (as created): a "world foundation model" that generates physically-plausible video of an environment's future states from text/image/video input, built to produce synthetic training data and simulation for physical AI (robots, autonomous vehicles).
Project Bernini¶
Creator: Autodesk Research (USA), 2024.
Input: multimodal prompts — e.g. text, images, 2D sketches, point clouds.
Output: generated 3D shapes / geometry.
Architecture (code) available? No
— a research prototype, not released.
Trained weights available? No
— described as "strictly experimental… not available for public use."
Purpose (as created): an experimental generative-3D foundation model, trained on roughly 10 million 3D shapes, that produces 3D geometry from multimodal input.
Stable Diffusion¶
Creator: Stability AI (UK) with CompVis (LMU Munich, Germany) and Runway (USA), 2022.
Input: a natural-language text prompt (plus an optional image for image-to-image or inpainting).
Output: a generated raster image.
Architecture (code) available? Yes
— open-source.
Trained weights available? Yes
— downloadable under the use-restricted CreativeML OpenRAIL-M licence.
Purpose (as created): a latent text-to-image diffusion model that generates images from text prompts by iteratively denoising in a compressed latent space.
-
Point-cloud encoders (OpenShape, ULIP-2, Uni3D). These embed 3D shapes sampled as point clouds — points scattered over a mesh surface — not as exact B-rep geometry. The study sidesteps this family for the CATPart-retrieval thesis (it prefers the B-rep/STEP-native models — UV-Net and its family), because a point cloud discards the exact CAD geometry. All three have open code and downloadable weights — fine to experiment with — but none carries a clean commercial licence; see each entry above for the specific terms. ↩↩↩