Start by loading the Krea 2 diffusion model in the checkpoint loader node. The tutorial begins with the NVFP4 model, which is only an 8GB download and works well for most users. For the CLIP model, the BF16 version is recommended for ComfyUI, but if you have lower system RAM, the FP8 option is a great alternative.

The model variants available include:

• NVFP4 (8GB) - smallest and fastest, recommended for low VRAM
• FP4 - good balance of quality and performance
• FP8 - for systems with lower RAM
• BF16 - highest quality, requires more VRAM

For loading LoRAs, use the RG3 Power LoRA Loader node which allows you to stack multiple LoRAs in one node.

The K sampler works with standard sampler/scheduler combinations. A great starting point is 8 steps at CFG 1. The model sampling node can help fix noisy artifacts and improve composition by adjusting the noise scheduling setting, which can transform generations for cleaner, sharper images.

For image-to-image work, adjust the denoise value based on how much you want to change the image. A denoise under 0.5 is usually the sweet spot for keeping at least some of the original image, though this depends on the sampler and scheduler used.

Conditioning rebalancing is a new feature described as "a conditioning optimizer with per-layer weighting that offers IP adapter-like features for ComfyUI, along with bypassing the built-in quality dilution from the trained safety filter."

In practice, it can improve prompt adherence but may reduce the model's ability to generate clear text. Set the mult value to around 0.5 for best results with most prompts. You can increase or decrease this value as needed.

The tutorial demonstrates that enabling conditioning rebalancing with a mult of 0.5 helped shrink an oversized portal in the generated image and improved adherence to the fantasy art style rather than photorealism.

You can load images and encode them along with your prompt using the text encode when image edit plus node. This acts as a reference image feature, even though Krea 2 isn't specifically an editing model.

The reference image strongly influences the output style and composition. In the tutorial, using the ComfyUI example image resulted in a more cartoony, childlike drawing style with features like clouds and distinctive hair - all without changing the prompt.

You can use up to three reference images simultaneously. When using multiple images, you may find they stack or form a grid - fix this by using a lower denoise value in the K sampler (around 0.3 with linear quadratic scheduler).

Krea 2 was trained on natural language and works fine with short, medium, or long prompts. The tutorial uses a complex prompt testing text, style, and positioning: "rodent mage summoning cheese using his biscuit tin."

The model handles various styles well:

• Watercolor - produces one of the better watercolor styles seen, correctly showing cake sponge inside a cake woman
• Cartoon - handles cartoon styles with correct text in speech bubbles
• Photorealistic - handles complex scenes with fire, ice, and wool buildings at 1920x1080

If you prefer AI assistance with prompts, use the built-in generate text node with the loaded CLIP model and a custom system prompt.

Krea 2 handles standard image-to-image generation by lowering the denoise in the K sampler. Use a denoise value under 0.5 to keep elements of the original image while applying prompt changes.

The tutorial demonstrates:

• Changing a character to an anime style with a Tokyo night market background using denoise 0.4
• Converting cartoon to photorealistic style with detailed skin textures using an empty latent
• Combining image encoding with image-to-image for more dramatic changes (can go up to full denoise when image is encoded)

For style changes, you can use an empty latent at any resolution while encoding the reference image, allowing you to change aspect ratios easily.

Krea 2 offers several upscaling options:

1. 2x VAE - Doubles resolution quickly without changing the image. The default generation is up to 2048x2048, and the 2x VAE can produce 4096x4096 results in seconds.

2. Ultimate SD Upscale - Keeps the image very similar to the original while adding detail. The tutorial shows a 1152x1536 result with clearer knitted wool effects.

3. Refiner stages - Using Krea as an upscaler through two stages of refining produces more photorealistic results. A 1200x1600 result featured clock eyes and more detail, though artifacts may appear (like on the neck at denoise 0.45).

The choice depends on whether you want minimal changes (2x VAE), slight detail enhancement (Ultimate SD Upscale), or more dramatic refinement (refiners).

You can load and encode up to three images for complex compositions. The tutorial demonstrates:

• Two images: a face and a pattern, producing a photorealistic portrait with the pattern as background
• Three images: a pattern, a rodent character, and a landscape, blended into one cohesive scene

For three images, use a lower denoise (around 0.3) with linear quadratic scheduler to prevent grid stacking. The result blended all three elements perfectly - the landscape with the rodent posing against a magical pattern sky.

Remember that Krea 2 isn't an editing model, so reference images act more like suggestions than strict guides.

When encoding images with empty latents, you can easily change the aspect ratio. The tutorial demonstrates going to 1080x1920 (portrait orientation) while asking for an oil painting style with thick brush strokes.

This works well for style transfers where you want to maintain the subject but change the composition orientation. Simply set the empty latent dimensions and encode your reference image.

For inpainting, use a masked latent. The tutorial uses SAM 3 segmentation to create masks automatically, though you can draw them yourself.

The example prompt asks for "the head of a bearded man" who "has perhaps seen better days." Krea 2 delivers exactly this, replacing only the masked area while keeping the rest of the image unchanged.

This provides a simple way to do head or face changes while preserving the rest of the composition.