Kreros 1.0 is a photorealistic merge built on Krea 2 Turbo (INT8 ConvRot), blended with amateur photography and realism LoRAs. It specializes in raw, candid, amateur-style photos — the kind of look you get from a smartphone shot in a bedroom, bathroom, or party: imperfect framing, natural window or flash lighting, real skin texture with blemishes and tan lines, and an overall unpolished, authentic feel. NSFW capable. Recommended settings: 8 steps, CFG 1.0, Euler / simple, 1K-2K resolution. Works great with vertical (9:16) prompts. Pair it with the standard Krea 2 text encoder (Qwen3VL 4B) and Qwen Image VAE. Tips: add words like 'candid', 'amateur photo', 'smartphone photo', 'shot on iPhone', 'flash photo at night' or 'analog film' to the prompt for stronger authenticity; describe lighting and camera angle explicitly for best results. Trained words: not required. This is an experimental personal merge shared for the community — no commercial restrictions on outputs. Base model license: Krea AI Community License. Merges and derivatives are allowed, credit is appreciated but not required.
Description
Kreros 1.0 NVFP4 — the ultra-lightweight, high-fidelity edition of Kreros 1.0 (Krea 2 Turbo with baked-in Realism and Identity Edit LoRAs), using ComfyUI's native NVFP4 (E2M1) format with 16-element FP8 (E4M3) micro-block scaling and FP32 tensor scaling. KEY HIGHLIGHTS: - Size drops from ~13.5 GB to only 8.17 GB (7.61 GiB) — fits effortlessly on 8GB, 12GB and 16GB VRAM GPUs. - Native NVFP4 (E2M1 + block FP8) preserves dynamic range far better than standard INT4. - Hardware acceleration: executed directly on 5th-gen Tensor Cores (Blackwell architecture). - Realism LoRAs fully baked into the weights (no separate LoRAs needed). TRADE-OFFS: retains ~95-98% of the INT8 visual fidelity; seeds are not bit-identical to the INT8 release. HARDWARE COMPATIBILITY: - BEST ON: NVIDIA RTX 50-Series (Blackwell — 5060 Ti, 5070, 5080, 5090) for native Tensor Core FP4 acceleration. - RTX 30/40 Series: works in modern ComfyUI via runtime dequantization, but the INT8 or W4A8 versions are recommended instead. HOW TO USE IN COMFYUI: - Load with the standard "Load Diffusion Model" (UNETLoader) node, weight_dtype: default. - CLIP/Text Encoder: qwen3vl_4b_fp8_scaled.safetensors (krea2 loader). - VAE: qwen_image_vae.safetensors. RECOMMENDED SETTINGS: Steps 8, CFG 1.0 (turbo), Sampler Euler, Scheduler Simple, Resolution 1024x1024 (or 832x1216 / 1216x832).