Open source and AI models
Effective October 12, 2026
Atelier is built on open-source software and openly licensed AI models. Its setup downloads them from their publishers onto your computer, where each stays under its own licence. Atelier only offers models whose licences allow commercial use. Licences checked against their publishers on 2026-10-12.
Built with DINOv3. Atelier's 3D tools use Meta's DINOv3 image encoder under the DINOv3 License, and its cutouts use Meta's SAM 3.1 under the SAM License. Both forbid use subject to ITAR or trade controls and for military, warfare, nuclear, espionage or weapons purposes (see the Acceptable Use Policy).
AI models
| Model | By | Licence | Used for |
|---|---|---|---|
| Z-Image Turbo | Tongyi-MAI (Alibaba) | Apache-2.0 | gen -m z-image (the default text-to-image model); fix's default engine; mflux on macOS |
| FLUX.2 [klein] 4B | Black Forest Labs | Apache-2.0 | gen -m klein-edit (the default image edit model in Atelier) |
| Qwen-Image-2512 | Qwen (Alibaba) | Apache-2.0 | gen -m qwen-2512 (text-to-image, strong at text in images) |
| Qwen-Image-Edit-2511 | Qwen (Alibaba) | Apache-2.0 | gen -m qwen-edit-2511 (image edit) |
| HiDream-O1-Image | HiDream.ai | MIT | gen -m hidream / hidream-edit (photoreal images and edits) |
| Gemma 4 E4B-it | Apache-2.0 | text encoder in ComfyUI's HiDream-O1 graph | |
| Ming-Image-0.1-Design | inclusionAI (Ant Group) | MIT | gen -m ming (posters, UI and layout design) |
| Pixal3D | Tencent ARC | MIT | to3d (image to 3D model; --views for multi-view) |
| TRELLIS.2 | Microsoft | MIT | to3d (shape and texture stages of the Pixal3D graph) |
| DINOv3 ViT-L image encoder | Meta | DINOv3 License | image encoder inside the Pixal3D / TRELLIS.2 3D graphs |
| MoGe 2 and MoGe 3 | Microsoft | MIT | camera field-of-view estimate inside to3d's Pixal3D graph |
| BiRefNet | Peng Zheng (ZhengPeng7) | MIT | cutout (default background removal) and the 3D graphs |
| SAM 3.1 | Meta | SAM License | cutout -p "..." (cut out by text prompt) |
| Depth Anything 3 Mono-Large | ByteDance Seed | Apache-2.0 | analyze (depth map of a reference) |
| Real-ESRGAN x4plus and x4plus anime 6B | Xintao Wang | BSD-3-Clause | restore / vectorize pre-upscale |
| 1xDeJPG OmniSR | Philip Hofmann (Helaman) | CC-BY-4.0 | restore (removes JPEG blocking before upscaling and vectorizing) |
| 4xBHI RealPLKSR and 4xBHI DAT2 photo upscalers | Philip Hofmann (Helaman) | CC-BY-4.0 | restore -m bhi / bhi-dat2 (4x photo upscaling that handles JPEG and WebP damage and noise) |
| SeedVR2 3B and 7B | ByteDance Seed | Apache-2.0 | upscale and enhance (images and video) |
| Marigold v2 albedo | Huawei Bayer Lab, EPFL, University of Bologna | Apache-2.0 | to3d --delight (removes baked-in lighting from textures) |
| FILM frame interpolation | Google Research | Apache-2.0 | video --interp (smoother motion) |
| Wan 2.2 (TI2V-5B, I2V-A14B) with lightx2v Lightning LoRAs | Wan-AI (Alibaba); lightx2v | Apache-2.0 | video -m wan / wan-fast (local video clips) |
| Qwen3-VL-8B-Instruct | Qwen (Alibaba) | Apache-2.0 | local critic (critique, auto) and image description |
| HPSv3 (NF4 build) | MizzenAI; NF4 conversion by stella221125 | Apache-2.0 | gen -n N --pick (ranks results by human preference) |
| DINOv2-base | Meta | Apache-2.0 | to3d fidelity score (compares a 3D model's renders with the reference) |
| MediaPipe Face Landmarker | Apache-2.0 | enhance --face (finds faces to restore; no identification) | |
| U-2-Net (u2net) via rembg | Xuebin Qin; rembg by Daniel Gatis | Apache-2.0 (weights); MIT (rembg) | analyze's subject mask and cutout --engine rembg |
| Krea 2 Turbo (with the darkbrush and style-reference LoRAs) | Krea | Krea 2 Community License v1 | gen -m krea / krea-style (Windows; installed by setup, used only after the user accepts Krea's licence) |
| Falconsai NSFW image detection (content filter) | Falconsai | Apache-2.0 | refkit/safety.py: checks every Krea 2 picture and blocks flagged ones (the filter Krea's licence requires) |
Credits: FLUX.2 [klein] 4B by Black Forest Labs; Built with DINOv3; Real-ESRGAN, Copyright (c) 2021 Xintao Wang; 1xDeJPG OmniSR by Philip Hofmann (Helaman), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); 4xBHI RealPLKSR and 4xBHI DAT2 by Philip Hofmann (Helaman), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing..
Programs setup installs
Downloaded from their publishers and run as separate, unmodified programs. Setup accepts their package licence agreements for you after you agree on its Setup screen.
| Program | Licence | Used for |
|---|---|---|
| Git | GPL-2.0-only | setup and updates |
| uv | MIT OR Apache-2.0 | Python environments |
| Python | PSF-2.0 | refkit and setup scripts |
| Node.js | MIT | svgo, gltf-transform, gltfpack |
| PowerShell 7 | MIT | setup |
| 7-Zip | LGPL-2.1-or-later (with BSD parts and the unRAR restriction) | unpacks ComfyUI (Windows) |
| Scoop | Unlicense OR MIT | installs the visual tools (Windows) |
| Homebrew | BSD-2-Clause | installs tools (macOS); Atelier turns its analytics off (HOMEBREW_NO_ANALYTICS) |
| Blender | GPL-3.0-or-later | 3D cleanup, turntables and views (Atelier's GPL scripts in studio/refkit/blender) |
| Blender MCP add-on and server 1.0.3 | GPL-3.0 | lets Claude Code drive Blender (only when the user adds Claude Code) |
| FFmpeg (gyan.dev build) | GPL-3.0 | video encoding and frames |
| ImageMagick | ImageMagick License | image conversion |
| ExifTool | Artistic-1.0-Perl OR GPL-1.0-or-later | metadata |
| oxipng | MIT | PNG optimisation |
| Tesseract OCR | Apache-2.0 | analyze (text in references) |
| potrace | GPL-2.0-or-later | vectorize |
| pngquant | GPL-3.0-or-later | PNG palette reduction |
| resvg | Apache-2.0 OR MIT | SVG rendering |
| libwebp (cwebp) | BSD-3-Clause | WebP export |
| libavif (avifenc) | BSD-2-Clause | AVIF export |
| Inkscape | GPL-3.0-or-later | SVG conversion |
| f3d | BSD-3-Clause | 3D previews |
| KTX-Software | Apache-2.0 | texture compression (Windows) |
| vtracer | MIT | vectorize |
| svgo | MIT | SVG optimisation |
| glTF-Transform CLI | MIT | 3D model optimisation |
| gltfpack (meshoptimizer) | MIT | 3D model compression |
| rembg | MIT | background removal without ComfyUI (always with -m u2net) |
| mflux | MIT | local image models on Apple Silicon |
| hpsv3-4bit | MIT | runs the HPSv3 ranker |
| Playwright and Chromium | Apache-2.0 (Chromium: BSD-3-Clause and others) | exports ComfyUI workflows |
| Claude Code | Anthropic Commercial Terms / Consumer Terms | optional; the user's own Claude plan or API key, billed by Anthropic |
Engine
| Component | Licence | Used for |
|---|---|---|
| ComfyUI (Windows portable NVIDIA build, with its embedded Python, PyTorch and packages) | GPL-3.0 | runs the image, video and 3D models |
| ComfyUI (macOS, optional engine) | GPL-3.0 | cutouts, depth and frame interpolation on Apple Silicon |
| ComfyUI workflow templates | MIT | the graphs refkit runs |
Online services
| Service | What for | Notes |
|---|---|---|
| Google Gemini API (Nano Banana images, Veo and Gemini Omni video) | optional, with the user's own API key; prompts and images go to Google | 18+; in the EEA, UK and Switzerland only the paid tier may be used in an app; images carry a SynthID watermark; Generative AI Prohibited Use Policy applies |
| Claude Code (Anthropic) | optional; the user signs in through Anthropic with their own plan or key | Atelier never handles Claude credentials |
| Supabase | Atelier accounts and subscription status (Cruxbit Labs' processor) | |
| Stripe | payments (Cruxbit Labs' processor) | |
| Resend | account emails (Cruxbit Labs' processor) | |
| Vercel | hosts cruxbitlabs.com (Cruxbit Labs' processor) | |
| Hugging Face, GitHub, Google Cloud Storage, winget, Scoop and Homebrew | where setup and first use download tools and models; they see the computer's IP address |
Parts of Atelier under other licences
| Files | Licence | Notes |
|---|---|---|
| Blender scripts (studio/refkit/blender) | GPL-3.0-or-later | Cruxbit Labs' scripts; GPL because they use Blender's Python API; source is the shipped .py files (studio/refkit/blender/LICENSE) |
| Workflow graphs (studio/workflows) | MIT | exported and modified from ComfyUI's workflow templates, Copyright (c) 2023-present Comfy Org |
| Fonts Outfit and Space Grotesk | OFL-1.1 | bundled in the app (npm @fontsource); licence text in the npm section |
| Login screen art (studio/app/src/account/art) | Cruxbit Labs | made with Z-Image Turbo (Apache-2.0); no third-party images |
Trademarks
Product names are used only to say what Atelier works with; they belong to their owners, who don't endorse Atelier: FLUX and Black Forest Labs; ComfyUI and Comfy (Comfy Org); Blender (Blender Foundation); Gemini, Nano Banana and Veo (Google); Claude and Anthropic; Krea; Qwen (Alibaba); Meta, SAM and DINOv3.
The full list, including every library built into the app and each licence's text, comes with Atelier: open Settings, Legal, Third-party licences.