Krea 2 ZHA Arch Style is a Krea 2 style LoRA, trained to produce a sculptural contemporary architectural language across exterior and interior spaces, including cultural, civic, transportation, infrastructure, residential, and public-space designs.
It emphasizes flowing shell-like forms, sharp faceted geometry, pale mineral surfaces, deep architectural voids, layered circulation, cantilevers, dark recessed glazing, ribbed cladding, and coherent asymmetrical massing. The LoRA is intended to create expressive but believable contemporary architecture rather than reproduce a specific building, landmark, or individual architect’s work.
Trigger word: zha_arch_style
Recommended strength: 0.5–1.0
General-purpose starting point: 0.9
A separate detailer LoRA may improve small-scale details and material definition, but it is optional and not required. If used, a strength of approximately 0.4 is a reasonable starting point.
Add zha_arch_style at the beginning of the prompt. For the most coherent results, describe the building type, functional circulation, viewpoint, materials, lighting, and site. The LoRA works best when expressive curved forms are balanced with believable structural logic.
Sample prompt:
zha_arch_style, a sprawling contemporary educational campus nestled within a lush urban park, captured from a wide-angle street-level perspective. The architecture features a series of interconnected sculptural volumes defined by sweeping white mineral concrete walls and faceted composite panels that flow seamlessly into undulating rooflines. A prominent central library serves as the focal point, characterized by a large, shell-like white stone facade with deep recessed openings and an expansive cantilevered terrace overlooking a public plaza. Connecting various academic pavilions are elevated pedestrian bridges and wide, terraced ramps that weave through the site, creating layered circulation paths. The building surfaces transition between smooth concrete and ribbed panels, punctuated by dark recessed glazing and large glass curtain walls that reflect the surrounding greenery. Soft, diffuse overcast daylight casts gentle shadows into the deep voids of the atriums, while small figures of students move along the walkways to provide a sense of scale. The campus sits harmoniously among mature trees and manicured lawns, with no signage or nearby buildings distracting from the coherent, sculptural massing of the institution.
In my testing, batch sizes of 2 or more sometimes produce complementary views of the same architectural concept. This is not guaranteed.
Training:
Dataset: 52 images, approximately 1.1 megapixels each
Captions: generated with GPT-5.6 Luna and edited manually
Training base: Krea 2 RAW
Recommended inference base: Krea 2 Turbo
Training software: OneTrainer
Training hardware: NVIDIA GeForce RTX 5070 Ti, 16 GB VRAM
Training result: 50 epochs, 1,250 steps
Optional prompt-generation workflow
The included ComfyUI workflows use Ollama and Gemma 4 12B to generate detailed architectural prompts for use with the zha_arch_style LoRA. The workflow is entirely optional; the LoRA can also be used with manually written prompts or any other language model.
Workflow v1.0 is optimized for conservative, structurally believable contemporary architecture. A two-stage sampler workflow is included for cleaner and more precise image outputs.
The workflow does not form part of the LoRA itself, it is included only as a prompt-testing aid.
Workflow notes:
The workflow is optional and is used to help generate architectural prompts.
It uses Ollama nodes with Gemma 4 12B. You can of course use any LLM model that you like.
The prompt output is intended for Krea 2 inference with the zha_arch_style LoRA.
The LoRA should still be tested independently with manually written prompts.
The workflow does not replace the LoRA or form part of the model itself, it´s just for testing prompts.
The two-stage sampler workflow produces cleaner and more precise image outputs.
The Ollama system prompt is included as a .txt document in the ZIP file alongside the optional workflows.
Ollama options settings:
num_ctx: 8192
repeat_last_n: 64
repeat_penalty: 1.12
temperature: 0.65
seed: fixed for repeatable testing
control after generate: fixed
top_k: 45
top_p: disabled
num_predict: 600
mirostat: disabled
min_p: disabled
tfs_z: disabled
Creative exploration options settings:
temperature: 0.80
seed: randomize
control after generate: randomize
License: This LoRA is a derivative of Krea 2 and is distributed subject to the Krea 2 Community License Agreement. It is not an official Krea product and is not endorsed by Krea. https://www.krea.ai/krea-2-licensing