#comfyui

INT8_ConvRot excels in both speed and image quality

Made these Liara & Feron gens back in May and somehow never shared them. Re‑discovering them now while replaying the Shadow Broker mission feels very on brand.

Read the blog post: https://www.duruoz.net/AI/2026/07/14/harry-furnis/ Introduction This analysis details the impact of varying LoRA weights applied to the ‘Harry Furnis’ style within a systematic testing framework. Utilizing DreamShaper SDXL Lightning as a base model, experiments were conducted in ComfyUI incrementing the LoRA weight from 0.1 to 2.0. This exploration investigates how the distinctive stylistic essence of Harry Furnis gradually influences the latent space of the AI image generation process, specifically concerning structural fidelity and prompt adherence. The goal is to provide insight into effective LoRA application for nuanced artistic control when producing AI illustrations with Stable Diffusion. Technical Analysis At a weight of 0.1, the ‘Harry Furnis’ LoRA introduces subtle textural changes to the base DreamShaper output. A noticeable increase in line-work density is observed within architectural details and character clothing, particularly visible on the building facades and figures' garments. However, the overall composition remains largely governed by the SDXL model's inherent aesthetic; shadows are softened and the chromaticity presents a balanced, modern palette. As weight increases to 0.5, the Furnis style begins to assert more dominance. Shadows deepen considerably, creating a pronounced contrast between illuminated areas and recesses. Architectural structures exhibit increasingly intricate detailing – timber framing gains complexity, and window patterns become sharper. A stylistic saturation emerges; while prompt adherence is maintained (e.g., depiction of a town scene under moonlight), the artistic direction leans demonstrably toward Furnis's illustrative aesthetic. The range between 1.0 and 2.0 represents significant stylistic bias. The image shows that structural integrity becomes subtly challenged as the LoRA attempts to fully imprint its visual vocabulary. While texture complexity peaks, a slight loss of photorealism can be observed. Details are exaggerated; the glowing embers intensify and become more prominent, influencing overall tonal range. The architectural hierarchy remains discernible but is rendered with heightened stylization, approaching a highly illustrative quality reminiscent of classic etchings. This demonstrates that beyond a certain threshold, increasing LoRA weight prioritizes stylistic expression over nuanced prompt interpretation within AI art generation. The BUDU AI platform's capacity to manage these weight tests allows for precise control and visualization of the effect. Artistic Conclusion The ‘Harry Furnis’ LoRA exhibits a strong visual signature that, even at low weights, subtly modifies DreamShaper's base style. Higher weights (1.5-2.0) deliver pronounced results – deeply saturated shadows, intense textural detail, and a clear departure from realism. This exploration highlights the critical balance between stylistic fidelity and structural coherence when employing LoRAs in AI Image Generation. The optimal weight for achieving artistic cohesion lies within the 0.4 - 1.0 range; weights beyond this point may introduce excessive bias, potentially compromising prompt adherence and overall image quality despite yielding visually striking results using SDXL models like DreamShaper. Further research into adaptive weight application based on compositional elements would unlock even greater potential for sophisticated AICreativity. #AIArt #ComfyUI #StableDiffusion #AIGeneratedArt #AIIllustration #AICreativity #AIImageGeneration #BUDUAI #LoRA #WeightTesting #SDXL #DreamShaper

話題の非検閲ローカル動画生成AI「Sulphur2」とは？導入方法まで徹底解説！LTX2.3派生のクオリティも嬉しい！

Imagine if Murderbot was hiding something far more intriguing under that helmet.

Flux2-klein-4B_clear_INT8_ConvRot_HQ