Cozy Creator

A fast Rust-based generative-media runtime; generate images and video on your local machine for free, or rent a powerful GPU by the hour. Private, uncensored. Cozy Creator is simpler and more performant than ComfyUI.

Install

Linux and macOS

curl -LsSf https://tensorhub.com/install.sh | sh

Homebrew (macOS and Linux)

brew install cozy-creator/tap/cozy

Windows (PowerShell)

irm https://tensorhub.com/install.ps1 | iex

Make a long-form video

MiniMax H3 turns a story into one continuous video with sound: your characters, places and dialogue, segment after segment. Write a story.yaml:

# story.yaml: three 10-second segments, one continuous shot, with sound
mode: turbo
style: Anime, hand-drawn cel shading, soft neon city light, one continuous camera move with no cuts.
references:
  - name: Hana
    kind: character
    image: ./hana.png        # your character art: background removed, described for you
  - name: Kumo
    kind: character
    description: A small AI companion, a translucent blue fox spirit made of light, with a bright, playful voice.
  - name: Studio
    kind: background
    description: A cluttered anime studio at night, sketches pinned to the walls, a big window over a neon Tokyo skyline.
  - name: Video 1
    kind: video
    video: ./twirl.mp4       # a short dance clip that guides Kumo's twirl
segments:
  - duration_s: 10
    prompt: |
      [Shot 1] The camera glides slowly toward <Hana>, hunched over her drawing tablet in <Studio>.
      <Kumo> flickers to life above the tablet. <Kumo> (S2), brightly, says: <d>[English] Still drawing? Let me help.</d>
  - duration_s: 10
    prompt: |
      [Shot 1] The same shot continues, circling the desk. <Kumo> twirls through the air like <Video 1>,
      and every line <Hana> draws lifts off the screen as glowing ink.
      <Hana> (S1) laughs and says: <d>[English] Okay, partner. Let's finish this episode together.</d>
  - duration_s: 10
    prompt: |
      [Shot 1] The same shot continues, rising past them as the glowing sketches swirl out the open window
      and become a dragon flying over the neon skyline. <Hana> and <Kumo> watch it, smiling.

Then render it on a rented GPU:

cozy login you@example.com
cozy run minimax-h3/long_form --input story.yaml --on new

Each segment is 5–15 s; references are shared across segments, and context carries the picture and sound from one segment to the next, so they play as one continuous shot. A photo is cleaned and described automatically; a video reference guides motion.

--on new rents a card fast enough for the model, at the lowest cost per run; it shows the price and asks before it rents. --on new:rtx-5090 picks the card.

Run on this computer

On Linux, Windows or macOS with a supported GPU: just run it; cozy starts the local machine when needed.

cozy run qwen-image-2/generate prompt='a koi pond at night under paper lanterns, mist on the water'

Rent a GPU, use it, end it

cozy machine rent
cozy machine rent rtx-5090
cozy run qwen-image-2/generate prompt='a paper boat on a rainy street' --on <name>
cozy machine end <name>

The first lists the GPUs and their prices; the second rents one and prints its name. A rental also ends itself after 15 minutes with nothing to run. cozy package ls lists the packages and cozy model ls the models.

Packages

Image

anima: Anima anime text-to-image package for Cozy
generate — models: anima

fibo-2: Cozy text-to-image and image editing with Bria Fibo 2.0, Dev and Edit Turbo
generate · generate_turbo — models: fibo-edit-vlm, fibo-2-dev, fibo-vlm, fibo-2-edit-turbo

flux-2: Cozy text-to-image and reference editing with FLUX.2 [dev], its Turbo LoRA, and FLUX.2 [klein] 9B and 4B
generate · generate_turbo — models: flux-2-dev, flux-2-dev-turbo

flux-3-klein: Cozy text-to-image and reference editing with FLUX.3 Image Klein, distilled and base
generate · generate_turbo — models: qwen3.5-9b-mtp, flux-3-klein-base, flux-3-klein-upsampler, flux-3-klein

hidream-o1: Cozy text-to-image and reference editing with HiDream-O1-Image, full and Dev
generate · generate_turbo — models: hidream-o1, hidream-o1-dev

hunyuan-image-3: Cozy text-to-image and reference editing with Tencent's HunyuanImage 3.0 Instruct and Instruct-Distil
generate · generate_turbo — models: hunyuan-image-3-instruct, hunyuan-image-3-instruct-distil

ideogram-4: Cozy text-to-image with Ideogram 4: JSON captions, layout boxes and rendered text
generate — models: ideogram-4, qwen3-vl-8b-instruct

krea-2: Cozy text-to-image with Krea 2: Raw, the full model, and Turbo, its 8-step distillation
generate · generate_turbo — models: krea-2, krea-2-turbo

moss-image: Cozy text-to-image with MOSS-Image (OpenMOSS): layout boxes for objects and rendered text
generate — models: moss-image

qwen-image: Cozy image generation and editing using Qwen-Image-2512 and Qwen-Image-Edit-2511, full and Lightning
generate · generate_turbo — models: qwen-image-2512, qwen-image-edit-2511, qwen-image-2512-lightning-lora, qwen-image-edit-2511-lightning-lora

qwen-image-2: Cozy image generation and reference editing using Qwen-Image-2.1
generate · generate_turbo · project-adapter-entity · project-adapter-setup · project-adapter-world_state — models: qwen-image-2.1, qwen-image-2.1-turbo

sdxl: Stable Diffusion XL text-to-image
generate — models: sdxl

Video

minimax-h3: MiniMax H3 video with sound: one continuous shot of any length, and its upscale
generate · upscale · project-adapter-shot · project-adapter-upscale — models: minimax-h3, minimax-h3-turbo-lora, open-h3-regenerate-2k, h3-latent-upscaler

Upscale

seedvr2: SeedVR2 7B temporally consistent video restoration
upscale — models: seedvr2-7b

Other tools

birefnet: BiRefNet background removal for supplied character references
remove_background — models: birefnet

qwen3.5: Qwen3.5: single-turn questions about text, images and videos
ask — models: qwen3.5-9b

qwen3.6: Qwen3.6: single-turn questions about text, images and videos
ask — models: qwen3.6-35b-a3b

Why Cozy

Source

cozy-creator/cozy-creator: the cozy binary and the TensorFS library.
cozy-creator/packages: the model packages.