Guide / June 2026
The local Krea 2 blueprint for ComfyUI.
Krea AI dropped Krea 2 as an open-weights image model. This guide shows the practical Comfy Desktop route: the local text-to-image workflow, the model download path, and the missing LoRA fix that can block the first run.
Release
June 22, 2026
Krea 2 open-weights release.
Turbo
8 steps
Fast text-to-image inference path.
RAW
LoRAs
Best checkpoint for training and research.
Fix
Warm Pastel
Manual LoRA download path included.
Model briefing
What Krea 2 changes for local image workflows.
Krea 2 is a 12B dense DiT image model built around aesthetic range and creative control. The release centers on two checkpoints that work together: RAW for the malleable base model and Turbo for the fast generation workflow most ComfyUI users will run first.
Training checkpoint
Krea 2 RAW
Use RAW when you want the malleable base checkpoint for fine-tuning, research, post-training, or LoRA work.
Fast inference
Krea 2 Turbo
Use Turbo for the local ComfyUI text-to-image workflow. It is the 8-step distilled checkpoint built for speed.
Accuracy note
Krea 2 has a community license and specific responsible-use terms. Review the official model cards and license before using it in a public product or commercial workflow.
Video walkthrough
Watch the install, then keep the checklist open.
The written guide below follows the same flow from the recording: Comfy Desktop, the Krea 2 template, the model downloads, and the manual Warm Pastel LoRA fix.
Install blueprint
The Comfy Desktop path that actually gets you running.
This is the clean local route from the transcript: a fresh Comfy Desktop instance, the Krea 2 text-to-image template, and one manual file fix if the downloader misses the LoRA.
Install or update Comfy Desktop
Create a clean Krea 2 instance
Search all templates for Krea 2
Choose the local text-to-image workflow
Download the model files
ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors
ComfyUI/models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
ComfyUI/models/vae/qwen_image_vae.safetensors
ComfyUI/models/loras/krea2_warmpastel.safetensorsFix the Warm Pastel LoRA if it fails
krea2_warmpastel.safetensors, use the manual Hugging Face download and place the file in the shared LoRA folder.C:\Users\USERNAME\ComfyUI-Shared\models\lorasRefresh Comfy and queue the prompt
Troubleshooting
The errors are usually missing files, not a failed install.
The first run can look broken while Comfy is still missing a model, LoRA, or a recent workflow node. Use this checklist before tearing the instance apart.
Missing LoRA
If Warm Pastel does not download, place it manually in the shared LoRA folder and refresh the workflow.
Template missing
Update Comfy Desktop or wait for the stable Desktop release if the template has not landed in your build yet.
API mix-up
The API workflow is different. For this guide, use the local Krea 2 text-to-image workflow in Comfy Desktop.
Hardware notes
Pick the file your machine can actually carry.
Krea 2 is not tiny. Start with the memory-friendly Comfy files when you are unsure, and treat the bigger BF16 weights as the higher-end path.
24GB VRAM
Comfortable local path
Matt saw roughly 15 seconds per generation on an RTX 4090 when not screen recording. This is the cleanest local experience.
16GB VRAM
Possible with the right file
Use lower-memory files such as FP8, keep resolution reasonable, and expect the exact ceiling to depend on your machine and settings.
Lower VRAM
Quantized experiments
MXFP8, NVFP4, and GGUF options can make experiments more approachable, but lower memory usually means more patience.
Quantized files such as FP8 and GGUF are there to reduce memory pressure. They can trade a small amount of fidelity for a much more practical local workflow.
FAQ
Quick answers before you queue another run.
What is Krea 2?
Krea 2 is Krea AI's foundation text-to-image model family. The open release includes RAW for training and research, plus Turbo for fast local text-to-image inference.
Should I use Krea 2 RAW or Krea 2 Turbo in ComfyUI?
Use Krea 2 Turbo for the practical local ComfyUI generation workflow. Use RAW when you are training LoRAs, fine-tuning, or doing research work.
Where do I put krea2_warmpastel.safetensors?
For Comfy Desktop's shared model setup, place the file in C:\Users\USERNAME\ComfyUI-Shared\models\loras.
Do I need the API workflow?
No. If you want to run Krea 2 locally, choose the Krea 2 text-to-image workflow from Comfy Desktop templates, not the API workflow.
Can Krea 2 run on a 16GB GPU?
It can be possible depending on resolution, model file, and quantization. The FP8 files are a better starting point for 16GB cards than the larger BF16 files.
Work with Matt
Building around open image models?
Sponsored videos, AI tool marketing, media buying, consulting hours, speaking, and group training. If your image tool needs practical creator attention, send the brief.