<h3 id="update-922-new:-int8-convrot-version">Update 9/22 New: INT8 ConvRot Version</h3>Fucktoy Noir KR2 is now available in a native INT8 ConvRot version (~12.4 GiB). This build uses architecture-aware INT8 quantization with ConvRot (group size 256) across the model's primary transformer layers, while leaving precision-sensitive components in BF16. ConvRot helps distribute weight outliers before INT8 quantization, allowing the model to retain much more of the original BF16 behavior and fidelity than straightforward INT8 compression.This version is particularly useful for RTX 20/30-series and other GPUs without native FP8 acceleration, where INT8 Tensor Core support can make it a better fit than the FP8 release. It also cuts the storage and memory footprint roughly in half compared with BF16. For RTX 40/50-series hardware, FP8 remains an excellent default; use INT8 ConvRot for efficient INT8 execution and broader hardware compatibility, or BF16 when maximum fidelity is the priorityFucktoy Noir KR2 is a custom NSFW Krea 2 Turbo checkpoint built around a balance of photographic realism, strong character rendering, cinematic contrast, and the darker, more stylized visual language suggested by the Noir name.The goal with KR2 is not simply to push one particular aesthetic. It is intended to retain the flexibility that makes Krea 2 useful while giving the model a more opinionated visual baseline: convincing skin and materials, strong lighting response, expressive faces, detailed environments, and compositions that hold together without requiring excessively complicated prompting.It is particularly well suited to photorealistic and cinematic image generation, but it remains responsive to changes in wardrobe, environment, lighting, camera position, framing, and overall art direction. It can move from clean contemporary photography into much moodier, higher-contrast imagery without immediately collapsing into a single baked-in look.This is a Krea 2 Turbo model, so it is designed around the fast Turbo sampling regime rather than long diffusion schedules. For our own testing, 8 steps / CFG 1 / Euler / Simple is the primary baseline and a good place to start.<h2 id="bf16-and-fp8-versions">BF16 and FP8 Versions</h2>Fucktoy Noir KR2 is provided in two precision levels. They contain the same underlying KR2 model, but they are stored and executed at different numerical precision.<h3 id="bf16-maximum-fidelity">BF16 — Maximum Fidelity</h3>The BF16 version (~23.9 GiB) is the reference version of KR2 and preserves the model weights at bfloat16 precision.Use BF16 when maximum image quality is the priority and your hardware can comfortably accommodate it. In side-by-side testing, BF16 retains slightly better micro-detail, fine texture and edge definition. The difference is not transformative: the overall composition, identity, aesthetic and behavior of the model remain very similar, but it can become visible when examining skin texture, hair, fabric, small environmental details and other high-frequency information.BF16 is therefore recommended for:<ul><li>Final-quality renders where hardware resources are not a concern</li><li>High-resolution generation</li><li>Detailed character and portrait work</li><li>Benchmarking or evaluating the model at its highest available precision</li><li>Users with sufficient VRAM/system memory who simply want the least quantized version available</li></ul>If you have the hardware and don't care about the additional storage and memory requirements, BF16 is the definitive version of KR2.<h3 id="fp8-recommended-for-most-users">FP8 — Recommended for Most Users</h3>The FP8 version (~12.2 GiB) is a purpose-built quantized release of the same checkpoint.Rather than simply forcing every tensor into FP8, KR2's FP8 release was constructed using a Krea 2-compatible mixed-precision quantization layout. Eligible model weights are stored as FP8 E4M3FN with freshly calculated scaling data, while the checkpoint maintains the structure required for proper Krea 2 execution in ComfyUI.This cuts the checkpoint size roughly in half while retaining remarkably close behavior to the BF16 reference.The principal compromise is a small reduction in micro-detail and sharpness. In our testing, FP8 can appear slightly softer than BF16 when the same generation is examined closely. This is an expected consequence of reducing weight precision rather than a different model or aesthetic.FP8 is recommended for:<ul><li>Most everyday generation</li><li>Lower VRAM overhead</li><li>Faster/easier model loading and switching</li><li>Systems where BF16 creates unnecessary memory pressure</li><li>Users maintaining large local checkpoint libraries</li><li>Iteration and experimentation where the small BF16 quality advantage isn't worth doubling the checkpoint footprint</li></ul>For the majority of users, FP8 should be the default download.<h2 id="which-version-should-i-use">Which Version Should I Use?</h2>If you aren't sure, start with FP8 or INT8 ConvRot versionIt gives you essentially the full KR2 experience at approximately half the checkpoint size and is the version we expect to make the most sense for normal generation.Move to BF16 when you have sufficient hardware and want to squeeze the maximum available fidelity from the checkpoint. BF16 becomes especially worthwhile when the output is intended as a final image rather than an exploratory generation, or when subtle texture and fine-detail reproduction matter.Importantly, these are not two differently trained editions of KR2. FP8 is a quantized representation of the BF16 reference checkpoint. Prompts and workflows developed against one should therefore translate naturally to the other.<h2 id="recommended-starting-settings">Recommended Starting Settings</h2>Architecture: Krea 2 Turbo Sampler: Euler Scheduler: Simple Steps: 8 CFG: 1.0 Denoise: 1.0 Text Encoder: Krea 2-compatible Qwen3-VL VAE: Qwen Image VAE

Files with SHA256: 8216797deb03b1d6068e9a332174a0f8cf3f88db05455efa7aaff15c64526a1f