Text-to-image generators can produce striking architecture, but they rarely produce the same building twice. For professional design workflows, where a rendering must track a specific reference image through repeated rounds of feedback, that inconsistency is more than an aesthetic nuisance; it can make generative tools unusable. A new open-access study published in Architectural Intelligence on 24 July 2026 tackles this problem directly, comparing three image-to-image workflows built on the FLUX.1 diffusion model to find out which one best preserves the visual content of a reference architectural rendering.
The research, led by Yang Yu and Ding Wen Bao with Yuqian Liu, evaluated three approaches for generating architectural aerial renderings. The first used prompt engineering alone, with the multimodal model Florence-2 examining a reference image and writing a detailed textual description that FLUX.1 then turned into a new image. The second fine-tuned FLUX.1 using LoRA, a low-rank adaptation technique, on a dataset of architectural renderings. The third used FLUX.1 REDUX, a companion model that conditions generation directly on the reference image itself, combining that visual signal with a Florence-2 caption. All three workflows were run under identical conditions so that the comparison isolated the effect of each consistency-enhancement mechanism.
The test material consisted of 100 high-resolution architectural aerial renderings drawn from the architecture platforms Gooood and ArchDaily. The images spanned clear, overcast and twilight lighting conditions, varied camera angles and distances, and a range of materials including glass walls, concrete surfaces and masonry facades. The researchers deliberately chose this diversity to broaden the applicability of the findings and to reduce the risk that LoRA fine-tuning would overfit to repetitive visual patterns. Eighty of the images were used for LoRA training, while twenty held-out images, never seen during training, served as the evaluation set.
Each workflow was tested on the same twenty evaluation images with standardized generation parameters: a maximum resolution of 1024 pixels, twenty inference steps, a guidance scale of 10, the uni_pc_bh2 sampler, and a random seed fixed per reference image across all three workflows. The experiments ran locally through ComfyUI on an NVIDIA RTX 3090 with 24 gigabytes of video memory and 32 gigabytes of system memory, using FLUX.1 in 16-bit floating point format. The authors note these hardware specifications as a reference point for future research, since loading FLUX.1 alone demands substantial memory.
To measure results, the team introduced a composite metric called the Difference-Amplified Weighted Score, or DAWS, which combines four established image-similarity measures: structural similarity (SSIM), pixel-level accuracy (MSE), reconstruction quality (PSNR), and a normalised histogram-distance score capturing colour-distribution consistency. The weights are heuristic, with SSIM weighted highest at 0.45 because architectural consistency depends primarily on spatial composition, structural relationships and geometric coherence. Squared terms amplify differences between stronger and weaker outputs, and the authors stress that DAWS is intended as a preliminary comparison framework rather than a replacement for expert architectural judgement.
The quantitative results told a nuanced story. On the composite DAWS measure, the image-conditioned REDUX workflow scored highest at 0.165, followed by prompt-only generation at 0.162, with the LoRA-enhanced workflow trailing at 0.147. REDUX also achieved the lowest mean squared error at 5897.137 and competitive PSNR at 10.539, marginally outperforming prompt-only generation. Interestingly, the prompt-only workflow achieved the highest SSIM at 0.204, just ahead of REDUX at 0.199, while the LoRA workflow scored lowest at 0.189. The authors interpret this as evidence that conventional pixel-level metrics do not fully align with architectural visual judgement, since SSIM values remained below 0.21 across all approaches and failed to capture semantic features such as building form and spatial relationships.
Visual analysis reinforced the quantitative picture. The prompt-only workflow generated aesthetically pleasing but highly variable outputs that failed to preserve specific design features, spatial compositions or lighting characteristics from the references. The LoRA-enhanced workflow showed marginal improvements in stylistic consistency, particularly in material palettes, but spatial accuracy and specific feature preservation remained limited. The REDUX workflow demonstrated the strongest overall reference-image fidelity, preserving architectural massing, facade articulation, rooftop configurations, lighting patterns and spatial relationships. The authors attribute this to image conditioning providing substantially stronger structural constraints than text-based approaches, particularly for complex spatial compositions and fine architectural details.
Perhaps the most striking finding concerns the LoRA fine-tuning itself. The process required more than 75,000 training steps and produced a 772-megabyte adapter, yet delivered only small improvements in colour consistency while performing worse on structural accuracy and overall visual quality than the unmodified model. The authors suggest that eighty images represent a small sample within the vast parameter space of FLUX.1, and that the memory demands of the adaptation process, which requires at least 24 gigabytes of GPU memory, further limit its practicality. They argue that when reference images are available, as is common in architecture and design, using them directly as conditioning signals may be a more practical and efficient path than expensive model retraining, challenging the common assumption that domain-specific fine-tuning is always necessary or optimal.
The study’s authors frame the image-conditioned workflow as offering practical advantages beyond raw fidelity: it eliminates the need for dataset-specific fine-tuning, enabling rapid deployment across diverse projects; it supports systematic collection of reference-output pairs as structured data assets for quality control; and by removing computational barriers it could make advanced generative tools accessible to resource-constrained practices. The work extends an earlier conference paper presented at CDRF, adding streamlined methodology and composite computational metrics built from conventional parameters.
The authors are candid about the study’s limitations. The research focuses on a single model family, FLUX.1, and the rapid pace of generative AI development means the findings will need validation across other models. The dataset of 100 renderings is modest, and future work should confirm the results with larger collections spanning broader architectural styles. The DAWS weighting scheme is heuristic and awaits calibration through correlation with human perceptual evaluation and expert judgement. Future directions include task-specific evaluation metrics that better capture architectural design intent, extended validation across diverse typologies and cultural contexts, and integration pathways for embedding the workflow into existing BIM and design software ecosystems. Within those bounds, the study offers a measured but useful signal: for architects seeking consistent AI-generated renderings, showing the model the reference image appears to work better than describing it.
Subject of Research: Comparing image control workflows in generative AI for architectural rendering consistency
Article Title: Comparative analysis of image control workflows in generative AI for architectural design: a FLUX.1-based study
Article References: Yu, Y., Bao, D. W., & Liu, Y. (2026). Comparative analysis of image control workflows in generative AI for architectural design: a FLUX.1-based study. Architectural Intelligence, 5(1), Article 23. https://doi.org/10.1007/s44223-026-00134-6
Image Credits: AI Generated
DOI: 10.1007/s44223-026-00134-6
Keywords: generative AI, architectural rendering, FLUX.1, image-to-image, LoRA fine-tuning, prompt engineering, FLUX.1 REDUX, image consistency, DAWS metric, diffusion models, ComfyUI, computational design
News Source: Denise Maddox. (October 28, 2026). Image-conditioned AI beats text prompts and fine-tuning at architectural rendering fidelity. Scienmag.
