What happened
Krea published the official inference code for Krea 2, an image model the repository says was trained from scratch and focused on creative and stylistic exploration. It offers two checkpoints, Krea 2 RAW and Krea 2 Turbo, alongside instructions for downloading the safetensor files and running inference. The repository identifies RAW as the base model: pretrained but not distilled, intended for fine-tuning, post-training and LoRA training. Turbo is an eight-step distilled checkpoint designed for fast text-to-image generation.
The two checkpoints are presented as parts of one workflow rather than alternatives. Krea recommends training LoRAs on RAW and applying them to Turbo for inference, saying LoRAs trained on RAW should work well on Turbo. Its example RAW command uses 52 sampling steps and classifier-free guidance (CFG) at 3.5. The Turbo example uses eight steps with CFG disabled and a constant timestep-shift value of 1.15. The README gives an output resolution range of roughly 1K to 2K for Turbo; it says the model was trained to generate up to 1K resolution for RAW. The general width and height options are listed as approximately 1024 to 2048 pixels.
The repository includes inference code and setup instructions using uv sync, environment variables for the downloaded RAW and Turbo checkpoints, and flags for prompts, steps, guidance, dimensions, seeds and image count. It lists ComfyUI, Fal and SGLang among platforms for running the open-source models. For fine-tuning, it names Hugging Face Diffusers, Ostris AI toolkit, Fal and Kohya (musubi tuner) as recommended providers or tools. These are repository instructions and recommendations, not a claim that each platform offers identical setup or performance.
The code repository is licensed under Apache-2.0, but the README says both model weights are under Krea’s community license with permissive use. It directs people seeking a commercial license to contact Krea. The repository reports 962 stars and 77 forks in the supplied source snapshot. It also cites an Artificial Analysis ranking, but the supplied material does not include the underlying evaluation results or enough detail to assess that comparison independently.
What the archive adds
This is a new Krea 2-specific implementation detail, while extending a related pattern of image-model distribution through visual workflow tooling. A prior note on day-one ComfyUI support with quantized variants covered a different model and adapter path; it does not establish that Krea 2 has the same formats. The earlier post shows that ComfyUI integration is a recurring adoption consideration, not that this is the same event.
Trend signal
The Krea repository supports the idea that generative media is becoming application infrastructure: it presents model checkpoints alongside inference code and routes into ComfyUI and other serving platforms. The signal is modest. The supplied material documents integration options, but does not establish adoption, production reliability or how the model performs in downstream applications.
Why it matters
For a team building an image workflow, RAW/Turbo separation makes the customization path more legible: train or adapt against a malleable base, then use the distilled checkpoint where a shorter sampling run matters. That is a practical design choice to test if a product needs a consistent house style, rather than assuming every inference call must use the same checkpoint used for adaptation.
The workflow also comes with operational decisions. ComfyUI support may make it easier to place Krea 2 in a visual pipeline, while the listed inference and fine-tuning tools point to several possible runtimes. Teams should validate memory requirements, output quality, speed and LoRA compatibility on their own workloads; the README examples alone do not settle those questions. Licensing needs a separate review: the Apache-2.0 label applies to the repository code, while commercial use of the model weights may require a different agreement.
Limits and uncertainty
The supplied source is a GitHub README and repository snapshot, not a full independent model evaluation. It does not provide hardware requirements, measured latency, throughput, VRAM use, training cost, benchmark methodology or a comparison against competing models. The cited Artificial Analysis rank is the repository’s characterization, and the material here does not include its supporting results. “Permissive use” is not a substitute for reviewing the community license, and the repository asks users seeking a commercial license to contact Krea. The material also does not establish whether every listed platform integration is maintained, equally capable or available under the same terms.
Skill to build now
Prototype the RAW-to-Turbo handoff with one small, clearly scoped style task. Run the supplied inference examples as a baseline, then train a LoRA on RAW using a supported tool and test it on Turbo. Record output consistency, generation time, hardware needs and any differences from running RAW directly. Before using results in a commercial product, review the model-weight license independently of the code license.
Reader takeaway
Krea 2’s noteworthy implementation choice is the paired workflow: RAW for adaptation, Turbo for eight-step generation, with ComfyUI among the documented routes to use it. Builders can evaluate that division now, but should treat performance and commercial rights as open questions rather than conclusions established by the repository.