Krea 2 Turbo with Vpipe

The bundled Krea workflow uses krea/Krea-2-Turbo with its original transformer, encoder and VAE.

Copy the generation files

Create ~/frok/pipelines/image/my-krea.local/ and place these public example files inside it:

• run.vpipeline
• meta.json
• VPIPE-LICENSE and NOTICE.txt — keep these with redistributed copies.

If installing beside the bundled Krea workflow, change the copied metadata id to vpipe:my-krea, give it a distinct name, and set default to false. Refresh Settings → Generation after saving.

If models are missing

Krea requires access to its gated Hugging Face repository. Sign in there and complete the access request first. Use a Hugging Face token from that account with permission to read the model.

In Frok Desktop → Settings → Advanced → API tokens, save the Hugging Face token, then quit and reopen Frok before using Prepare models. For browser mode or direct Vpipe commands, supply HF_TOKEN in the launch environment. Vpipe reads this variable when the starter's hf_token field is empty; leave secrets out of pipeline files. An HTTP 401 or 403 means the download was denied: check the token, model access, and token permissions before retrying.

If Frok reports missing models for this workflow, copy the preparation pipeline below into a file named prepare.vpipeline, or download it. This is the source-controlled preparation file supplied with Frok:

{
  "id": "prepare-krea-2-turbo",
  "description": "Download Krea 2 Turbo, including its transformer, encoder, tokenizer and VAE.",
  "stages": [
    {
      "id": "fetch-krea-2",
      "type": "model-fetch",
      "iports": [],
      "config": {
        "model_path": "krea/Krea-2-Turbo",
        "hf_token": "",
        "source": "",
        "base_path": "./models",
        "skip_existing_files": true,
        "overwrite_existing": false
      }
    }
  ],
  "subpipelines": []
}

Keep the file in a location of your choice and run it from your configured Vpipe workspace:

cd ~/vpipe
vpipe --launch /absolute/path/to/prepare.vpipeline

Use your configured workspace if it differs from ~/vpipe. After preparation finishes, choose Settings → Generation → Refresh workflows. Ready workflows do not need this file. For the unchanged bundled workflow, Prepare models in Frok queues this same script. It is kept with the app, outside the editable pipeline directory.

Preparation fetches the base model into the Vpipe workspace's models/krea/Krea-2-Turbo layout, reusing existing files. Generation selects that model, encodes the prompt, runs diffusion, decodes the final latent and saves an image. The supplied native example starts at 1024 × 1024, eight steps and seed 42; Frok overrides bound dimensions, prompt, seed and output location for each job.

These download files are snapshots of the included factory pipeline, not an experimental custom fusion profile.

Run directly in Vpipe

Once the model is installed, you can also run the generation file directly from the same Vpipe workspace without Frok:

vpipe --launch /absolute/path/to/run.vpipeline

Edit the native text-prompt stage and output filename before subsequent direct runs; the example output may be overwritten.

Provider access may require accepting a model license or supplying credentials through your runner configuration. Do not put tokens in a distributable pipeline file. Direct Vpipe jobs are outside Frok's queue, so avoid running them alongside Frok's GPU work.

Change a workflow, not a global model

To adjust sampling, edit the scheduler/generator settings in your copy. To add a compatible LoRA, configure it in the native graph and declare its required files in metadata. A fused variant needs a preparation graph that produces a separate model artifact; simply adding a metadata entry does not fuse anything.

Preparation and verification explains the difference between downloads and generated artifacts.