Before using Krea 2, you need to update ComfyUI and install the model files in the correct folders.

1. Update ComfyUI – run the latest ComfyUI update so the node system recognizes the new model.
2. Update Pixeroma nodes – use Easy Install to get the newest version of the Pixeroma custom nodes.
3. Clear the ComfyUI cache – press Ctrl+Shift+R to ensure all nodes load correctly.
4. Download the Krea 2 FP8 model (recommended for quality/speed balance):
   • Click the download button inside the Pixeroma node or grab the files from pixeroma.com
   • Place the model in ComfyUI/models/diffusion_models/ (create a Krea2 folder to stay organised)
5. Download the Qwen text encoder:
   • Place it in ComfyUI/models/text_encoders/
6. Download the Qwen image VAE:
   • Place it in ComfyUI/models/vae/

This is the basic workflow – just a prompt and a K‑sampler. You can download the ready‑made workflow from pixeroma.com or Discord.

1. Load the workflow in ComfyUI (download links on pixeroma.com or Discord).
2. Check the Pixeroma node – it shows the model (FP8, 12 GB). If you see “missing model”, press R to refresh the node definition and select the model from the dropdown.
3. Connect the Krea 2 enhancer node (optional but recommended):
   • This custom node improves prompt understanding
   • Set strength between 0.1 and 1.0 – the creator prefers 0.1
4. Set the K‑sampler parameters:
   • Steps: 8
   • CFG: 1 (no negative prompts needed)
   • Sampler name: ERSTE (Euler also works)
   • Scheduler: simple
5. Define the image size using the resolution node (quick ratio/width/height selection) or manually set width and height in the Empty Latent node.
6. Seed – use a fixed seed for reproducibility or switch to random for varied results.

With everything connected, you’re ready to generate.

1. Click Queue Prompt.
2. Watch the timer – a 12‑GB VRAM card generates a 1024×1024 image in about 15 seconds.
3. To get a new image each time, change the seed to random (right‑click seed input → set to random).

LoRAs trained on Krea 2 are already available on CivitAI. This workflow shows how to apply one.

1. Download a Krea 2 LoRA (e.g., “vintage tarot style” from the ComfyUI team) and place it in ComfyUI/models/loras/Krea2/.
2. Add the Power LoRA Loader node (from RG3 node pack – included in Easy Install).
3. Connect it between the Krea 2 model node and the Krea 2 enhancer node.
4. Select your LoRA and set its strength.
5. Add a trigger word to your prompt (e.g., “vintage tarot style”).
6. Run the workflow – it will take slightly longer but should produce the desired style.

Krea 2’s text encoder can turn a short phrase into a rich, detailed prompt.

1. Load the prompt‑enhanced workflow (available from the same download links).
2. Connect a Text Generate node to the clip text encoder.
3. Enter a formula (e.g., a ChatGPT‑generated template) and a short prompt like “cute bunny”.
4. Run the workflow – the Show Text node displays the expanded prompt that will be used.
5. The image is generated from that expanded prompt.

If you have 6 GB of VRAM, use the low‑VRAM variant of the workflow.

1. Load the low‑VRAM workflow (included in the downloads).
2. The main difference is a Tiled VAE Decode node – this helps avoid crashes on limited VRAM.
3. Use smaller image sizes first (e.g., 512×512) to test stability. Larger images require more VRAM.
4. Generation time on a 6 GB card with 64 GB system RAM was under one minute.

This two‑step method generates a base image, then upscales and refines it for richer detail.

1. Load the extra‑pass workflow.
2. After the first K‑sampler, a Latent Upscale node scales the image by 1.5× (higher values may break the image).
3. A second K‑sampler runs with only 4 steps and denoise 0.4–0.5.
4. The result is a larger, more detailed image.

Combine the extra‑pass technique with a LoRA or the prompt enhancer.

• With LoRA – connect a Power LoRA Loader to both K‑samplers so the style is applied consistently across both passes.
• With prompt enhancer – add the same text‑generation chain used in Step 5; the expanded prompt feeds into both samplers.

This is the workflow used to create images for 2K YouTube thumbnails and screenshots.

1. Load the 2K workflow.
2. Choose portrait or landscape – width and height are pre‑set so that after a 1.5× upscale you get a 2K image.
3. Optionally, add the prompt enhancer for automatic detailed prompts.
4. Run the workflow – on a 24 GB card, generation takes about 40–42 seconds.
5. Inspect the result – textures, clothing details, and overall quality are remarkably high.

Krea 2 handles a wide range of styles. Here are some results from the creator’s tests (all using the 2K extra‑pass workflow with prompt enhancer):

| Prompt (short) | Result | |----------------|--------| | “Macro photo of a burger” | Realistic burger, add camera terms for more photorealism | | “Game interface with buttons” | Clean UI, occasional text errors (e.g., “Settngs”) – regenerate seed to fix | | “Anime style” | Good anime quality | | “3D style anime” | 3D look, still anime | | “Pencil drawing of an old witch” | Slight plastic look – adjust prompt to avoid “dust particles” | | “Watercolor painting of a unicorn” | Excellent paper texture | | “Modern building at golden hour” | Architectural rendering look | | “Lamb and frog holding an umbrella” | Can struggle with object interaction – re‑seed or simplify | | “Cinematic title ‘The End’” | Cinematic look |