<h2 id="krea2-sat-ior-imitation-of-reality">Krea2-SAT-IOR Imitation Of Reality</h2>IMPORTANT NOTE: Here are the GGUF Q4_K model and the NVFP4 model. Civitai detects the NVFP4 as BF16 quantization and this cannot be changed; it is not BF16.<span style="background-color:rgb(37, 38, 43);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:14px">The INT4 ConvRot SR file was also added.INT4 ConvRot SR – Performance Edition This quantization is primarily intended for NVIDIA GPUs that can efficiently accelerate low-precision INT4 workloads. Performance gains depend heavily on your GPU architecture, inference backend, and available kernels.On my RTX 4090 (Ada Lovelace), at 1152×1440 resolution, this version reduced generation time from approximately 11 seconds with FP8 to ~4.7 seconds, using the same generation settings.The model size is also significantly reduced, from roughly 12 GB with FP8 to just under 7 GB, making it considerably lighter in terms of storage and VRAM requirements.INT4 ConvRot SR trades some numerical precision for substantially lower memory bandwidth and much faster inference when the hardware/backend can take advantage of it. A small loss of fine detail may become noticeable compared with the original FP8 model, so I recommend FP8 for maximum quality and INT4 ConvRot SR for maximum speed.Recommended: modern NVIDIA RTX GPUs with good INT4/low-precision acceleration, especially when supported by optimized inference kernels. Results may vary significantly between GPU generations and software backends.The INT8 and CONVROT quantizations are now up.IOR v2 — GGUF Q4_K EditionIOR v2 is now available in GGUF Q4_K, bringing the same model to users with more limited VRAM without requiring the full FP8 checkpoint.This version was quantized directly from the original IOR v2 model. Despite the much more aggressive quantization, my testing showed surprisingly little visual degradation. The characteristic realism, natural skin detail, lighting, textures and overall photographic behavior of IOR v2 remain remarkably close to the original.The Q4_K version offers a significantly lighter alternative for users with less VRAM.FP8 remains the recommended version if you have enough VRAM and want the original checkpoint at its highest available fidelity. GGUF Q4_K is intended as the more accessible option.Same IOR. Less VRAM.NVFP4 VersionNVFP4 is NVIDIA’s 4-bit floating-point quantization format, designed primarily for Blackwell-generation GPUs such as the RTX 50 series, where native FP4 hardware acceleration can provide the greatest benefit.This version significantly reduces model size and VRAM requirements while preserving most of the original model quality. It may also run on older NVIDIA GPUs through supported inference implementations, although they do not have Blackwell’s native FP4 acceleration.This IOR v2 NVFP4 was converted directly from the original checkpoint. In my testing it produced valid generations at essentially FP8-like speed, despite the substantially smaller model size.Recommended for: RTX 50-series / Blackwell GPUs and users looking for a lower-VRAM SafeTensors version.For maximum fidelity: use the original FP8 version.Recommended Settings<ul><li>Width and height: 896*1152 base and 1152x1440 Hires</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>The sampling method i use is Euler-SGM Uniform, You can also try ER SDE-SGM Uniform for a bit more rawness.</li></ul><h2 id="about-this-version">About this version</h2>IOR V2 — Imitation of RealityIOR V2 is a major evolution of the original IOR for Krea2, built around one simple idea: reality is imperfect.This version pushes the model further away from the polished, overly controlled look commonly associated with AI-generated imagery and toward something that feels photographed rather than generated.IOR V2 improves natural skin and material texture, believable human expressions, complex lighting, photographic depth, environmental coherence and fine detail, while preserving Krea2's excellent prompt understanding and versatility.But perhaps the most important change is not simply image quality. It is how IOR V2 thinks about a photograph.Sometimes the best way to describe its behavior is:“Who the hell took this photo?”IOR V2 is comfortable with awkward framing, partially obscured subjects, objects intruding into the foreground, strange camera positions, harsh flash, imperfect exposure and moments that don't look carefully staged. Instead of constantly trying to create the most beautiful composition possible, it can produce images that feel like someone was actually there and happened to press the shutter.This also helps impossible subjects feel strangely believable. A creature, historical scene or surreal event doesn't necessarily need to look cinematic. Sometimes it becomes more convincing when it looks like an ordinary photograph of something that absolutely should not be there.How is IOR V2 different from Dirty Realism?Although they share some DNA, IOR V2 is not Dirty Realism V5.Dirty Realism aggressively pursues rawness, texture and photographic grit. It has a stronger aesthetic fingerprint and intentionally pushes Krea2 toward a particular kind of imperfect realism.IOR V2 takes a broader approach.Rather than imposing one specific photographic aesthetic, it focuses on believability, photographic behavior and versatility. It can be clean or dirty, modern or historical, beautiful or uncomfortable, documentary or fantastic, while still trying to preserve the little inconsistencies that make a photograph feel real.Dirty Realism has a look. IOR V2 has a photographic instinct.IOR V2 isn't trying to create a perfect reality.It's trying to imitate the imperfect one we actually photograph.Recommended Settings<ul><li>Width and height: 896*1152 base nad 1152x1440 Hires</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>The sampling method i use is Euler-SGM Uniform, You can also try ER SDE-SGM Uniform for a bit more rawness.</li></ul>Welcome to the first release of the IOR (Imitation of Reality) series.While Dirty Realism was designed to capture the raw, imperfect beauty of real-world photography, IOR takes a different path. Its mission is simple: create images so clean, detailed, and naturally balanced that they become almost indistinguishable from professional photographs.IOR Krea2 V1 has been carefully optimized to maximize resolution, fine detail, and image fidelity. Every aspect of the model has been refined to produce sharper textures, cleaner edges, realistic lighting, and highly accurate materials without sacrificing the natural look that makes an image believable.The result is a model capable of generating photographs with exceptional clarity while preserving authentic colors, realistic skin, physically convincing lighting, and true-to-life environmental details.<h2 id="why-you'll-love-ior">Why You'll Love IOR</h2><ul><li>Exceptional high-resolution performance.</li><li>Outstanding detail retention, even in complex scenes.</li><li>Clean, realistic images with minimal artifacts.</li><li>Highly accurate lighting, materials, and textures.</li><li>Natural skin tones and lifelike facial details.</li><li>Excellent consistency across a wide variety of prompts.</li><li>Optimized for photorealism without looking overprocessed.</li><li>Perfect for portraits, architecture, interiors, products, fashion, landscapes, automotive, and commercial-quality photography.</li></ul><h2 id="the-philosophy">The Philosophy</h2>IOR stands for Imitation of Reality, and that is exactly what this model was built to achieve.Instead of emphasizing stylization or dramatic effects, IOR focuses on reproducing the subtle visual characteristics that make a photograph feel genuine. Every generation aims to recreate the balance, clarity, depth, and precision expected from high-end professional cameras and modern photography.Whether you're creating commercial renders, architectural visualizations, realistic portraits, lifestyle photography, or cinematic environments, IOR Krea2 V1 is designed to deliver clean, high-fidelity images that look ready to publish.If you're looking for one of the most realistic Krea2 models available, give IOR Krea2 V1 a try. Download it, push it to its limits, and see how close AI can come to reality.<h2 id="recommended-settings">Recommended Settings</h2><ul><li>Width and height: 896*1152 base nad 1152x1440 Hires</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>The sampling method i use is ER SDE-Simple</li></ul>

Krea2-SAT-IOR (Imitation of Reality)

IOR v2 — GGUF Q4_K EditionIOR v2 is now available in GGUF Q4_K, bringing the same model to users with more limited VRAM without requiring the full FP8 checkpoint.This version was quantized directly from the original IOR v2 model. Despite the much more aggressive quantization, my testing showed surprisingly little visual degradation. The characteristic realism, natural skin detail, lighting, textures and overall photographic behavior of IOR v2 remain remarkably close to the original.The Q4_K version offers a significantly lighter alternative for users with less VRAM.FP8 remains the recommended version if you have enough VRAM and want the original checkpoint at its highest available fidelity. GGUF Q4_K is intended as the more accessible option.Same IOR. Less VRAM.NVFP4 VersionNVFP4 is NVIDIA’s 4-bit floating-point quantization format, designed primarily for Blackwell-generation GPUs such as the RTX 50 series, where native FP4 hardware acceleration can provide the greatest benefit.This version significantly reduces model size and VRAM requirements while preserving most of the original model quality. It may also run on older NVIDIA GPUs through supported inference implementations, although they do not have Blackwell’s native FP4 acceleration.This IOR v2 NVFP4 was converted directly from the original checkpoint. In my testing it produced valid generations at essentially FP8-like speed, despite the substantially smaller model size.Recommended for: RTX 50-series / Blackwell GPUs and users looking for a lower-VRAM SafeTensors version.For maximum fidelity: use the original FP8 version.Recommended Settings<ul><li>Width and height: 896*1152 base and 1152x1440 Hires</li><li>CFG Scale: 1</li><li>Steps: 8</li><li>The sampling method i use is Euler-SGM Uniform, You can also try ER SDE-SGM Uniform for a bit more rawness.</li></ul>

Files with SHA256: 578136f23cdf54d1696f0f12b62054347cd5065214245faf0587fd3676274b7e