> ## Documentation Index
> Fetch the complete documentation index at: https://docs.voyager.so/llms.txt
> Use this file to discover all available pages before exploring further.

# Machine images and versions

> Which software each Voyager cloud machine image carries, how to pin a version so a project keeps it, and what changed in every version.

A machine starts from an image: the operating system and the tools on it. Images come in families, and each version has a name like `creative@17`.

## Name an image

`image` on `POST /v1/machines` (and `--image` on `voyager cloud lease`) takes:

| You send | You get |
| - | - |
| nothing | the preset's image: the `creative` family's current version |
| `creative` or `creative@latest` | the family's current version |
| `creative@17` (or `creative@v17`) | exactly that version, until you change it |
| `creative-cuda@v17` | the same image by its registry name |
| `comfyui@1` or `ml@1` | version 1 of that family: ComfyUI with its models (open it on port 8188), or the PyTorch training stack |
| `creative-view` or `creative-view@1` | creative\@17 plus a desktop for [`voyager cloud view`](https://docs.voyager.so/api/machines/blender#blenders-ui-in-your-browser): Blender's UI in your browser |

The machine's `image` is the registry name it runs, and `image_pin` is the name that keeps it (`creative@17`). An unknown name is refused with `400 invalid_request`.

## List images and versions

`GET /v1/machines/images` lists each family's current version. Add `?all=true` for every version; older ones are marked `hidden: true` and you can still lease them. Each image has `pin`, `family`, `version`, `released` (YYYY-MM-DD) and `changes` (what changed from the version before).

## When the default moves

Voyager announces a default move before it happens. While one is announced, the family's images in `GET /v1/machines/images` carry `next_default` (`{name, on}`), and every machine on the current default carries `image_notice`, which names the pin that keeps it. Lease with that pin to stay on your version.

## Pin a project

* `voyager cloud images` lists each family's current version, with what changed; `--all` adds older ones.
* `voyager cloud images pin [name]` saves a version (default: the current `creative`) in `voyager-cloud.lock.json` in your project. `lease` and `run` then use it unless `--image` names another. Commit the file with your project.
* `voyager cloud lease` prints the default-move notice when there is one.

## Changelog

Generated from Voyager's image registry.

### Creative (`creative`)

Blender LTS and latest, Vork, Vauss, ffmpeg with NVENC, PyTorch, the Voyager assets core library and the Voyager cloud CLI: the default image.

Current version: `creative@18`.

| Version | Released | What changed |
| - | - | - |
| `creative@18` (current) | 2026-11-18 | The default creative image (`creative@17` keeps v17). Over v17: voyager-cloud 0.5.0, also as `voyager` on PATH (`voyager cloud …` works on the machine); `blender` runs Blender 5.3 for scenes saved by 5.3 and 5.2.4 LTS otherwise, saying which on stderr. |
| `creative@17` | 2026-11-14 | Vauss 0.1.23-rc (COLMAP-triangulated seeding). |
| `creative@16` | 2026-11-13 | Vauss 0.1.22-rc and supervisor memory fixes. |
| `creative@15` | 2026-11-12 | Vauss 0.1.21-rc. |
| `creative@14` | 2026-11-12 | Vauss 0.1.20-rc (security update). |
| `creative@13` | 2026-11-12 | Vauss 0.1.19-rc, public and pinned, replaces the private splat engine. |
| `creative@12` | 2026-11-11 | Blender renders on the GPU by default (OptiX, then CUDA; a CPU fallback warns); GPU use reports; voyager-cloud 0.4.0. |
| `creative@11` | 2026-11-11 | Blender 5.2.4 LTS as `blender` and Blender 5.3.0 beside it as `blender-5.3`; voyager-cloud 0.3.8. |
| `creative@10` | 2026-11-11 | Public Vork 0.5.1 with its baked shader cache. |
| `creative@9` | 2026-11-10 | Node.js 26 LTS with npm, NumPy and Pillow for every python3, voyager-cloud 0.3.4. |
| `creative@8` | 2026-11-09 | Voyager assets core library for offline use; voyager-cloud 0.3.1. |
| `creative@7` | 2026-11-09 | Vork 0.3.0 with its shader cache. |
| `creative@6` | 2026-11-09 | Vork 0.2.1 with its shader cache. |
| `creative@5` | 2026-11-07 | Multi-GPU Vork (`--gpu N`) and a baked shader cache. |
| `creative@4` | 2026-11-07 | Sets XDG\_RUNTIME\_DIR (no libwayland warning). |
| `creative@3` | 2026-11-06 | Vork 0.1.0. |
| `creative@2` | 2026-10-14 | Blender 5.2.2 (Cycles OptiX/CUDA, EEVEE) and the Voyager splat engine. |
| `creative@1` | 2026-10-14 | First image: PyTorch (CUDA 12.8), NumPy, Pillow, Warp, moderngl, ffmpeg with NVENC, NVIDIA EGL. |

### ComfyUI (`comfyui`)

ComfyUI with a curated node set and permissively licensed models on local disk (FLUX.1 \[schnell], Wan 2.2 TI2V 5B); it also reads models from a volume at /persist/comfyui/models, the workspace's model cache.

Current version: `comfyui@1`.

| Version | Released | What changed |
| - | - | - |
| `comfyui@1` (current) | 2026-11-18 | First image: ComfyUI 0.45.0, five curated nodes, FLUX.1 \[schnell] fp8 and Wan 2.2 TI2V 5B. Open to all workspaces. |

What it bakes in, with each licence:

* ComfyUI (GPL-3.0, source: [https://github.com/comfyanonymous/ComfyUI/tree/989d93d0d968e704c589ce0146bb18c77f6b50cc](https://github.com/comfyanonymous/ComfyUI/tree/989d93d0d968e704c589ce0146bb18c77f6b50cc))
* ComfyUI-Manager (GPL-3.0, source: [https://github.com/Comfy-Org/ComfyUI-Manager/tree/3a2177fb2484f086e554145add56dd282a7c5ad4](https://github.com/Comfy-Org/ComfyUI-Manager/tree/3a2177fb2484f086e554145add56dd282a7c5ad4))
* ComfyUI-VideoHelperSuite (GPL-3.0, source: [https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/tree/4d907bee61e92c2e65af3bd6383a4e4d356126d1](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/tree/4d907bee61e92c2e65af3bd6383a4e4d356126d1))
* comfyui\_controlnet\_aux (Apache-2.0, source: [https://github.com/Fannovel16/comfyui\_controlnet\_aux/tree/c2655d2970d810d8a3d71deed204cdf416b66132](https://github.com/Fannovel16/comfyui_controlnet_aux/tree/c2655d2970d810d8a3d71deed204cdf416b66132))
* rgthree-comfy (MIT, source: [https://github.com/rgthree/rgthree-comfy/tree/bc48c6aaeca87bedbc14245d89f9bc6be37a630e](https://github.com/rgthree/rgthree-comfy/tree/bc48c6aaeca87bedbc14245d89f9bc6be37a630e))
* ComfyUI-KJNodes (GPL-3.0, source: [https://github.com/kijai/ComfyUI-KJNodes/tree/4d6f027c23991aef33698a5c8c57452111df4cdb](https://github.com/kijai/ComfyUI-KJNodes/tree/4d6f027c23991aef33698a5c8c57452111df4cdb))
* flux1-schnell-fp8.safetensors (Apache-2.0, [https://huggingface.co/Comfy-Org/flux1-schnell/blob/d31e04067066d92c55ece2b01748dcc337725d7a/flux1-schnell-fp8.safetensors](https://huggingface.co/Comfy-Org/flux1-schnell/blob/d31e04067066d92c55ece2b01748dcc337725d7a/flux1-schnell-fp8.safetensors))
* wan2.2\_ti2v\_5B\_fp16.safetensors (Apache-2.0, [https://huggingface.co/Comfy-Org/Wan\_2.2\_ComfyUI\_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split\_files/diffusion\_models/wan2.2\_ti2v\_5B\_fp16.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors))
* umt5\_xxl\_fp8\_e4m3fn\_scaled.safetensors (Apache-2.0, [https://huggingface.co/Comfy-Org/Wan\_2.2\_ComfyUI\_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split\_files/text\_encoders/umt5\_xxl\_fp8\_e4m3fn\_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors))
* wan2.2\_vae.safetensors (Apache-2.0, [https://huggingface.co/Comfy-Org/Wan\_2.2\_ComfyUI\_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split\_files/vae/wan2.2\_vae.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/ee6f4a40737a995bf5818954cfce6d59443b0f04/split_files/vae/wan2.2_vae.safetensors))

### ML training (`ml`)

PyTorch 2.8 (CUDA 12.8) with Transformers, Accelerate, PEFT, Datasets, Lightning, xformers and JupyterLab, for training and fine-tuning; checkpoints belong on a volume.

Current version: `ml@1`.

| Version | Released | What changed |
| - | - | - |
| `ml@1` (current) | 2026-11-18 | First image: PyTorch 2.8 (CUDA 12.8) training stack. Open to all workspaces. |

### Base (`base`)

Python, ffmpeg and the NVIDIA libraries only, for smoke tests.

Current version: `base@1`.

| Version | Released | What changed |
| - | - | - |
| `base@1` (current) | 2026-10-14 | Python 3.12, ffmpeg, NVIDIA EGL libraries. |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.