krea2_raw_bf16

[NOTE: This is a standalone, unquantized Krea 2 RAW DiT model, uploaded under the FLUX.1 category due to architectural similarity and platform limitations.]

Official release of the Krea 2 RAW base model in native bfloat16 (BF16) precision.

Unlike the distilled Turbo variant, this RAW model captures the full mathematical accuracy of the 12B parameter DiT architecture from Krea.ai. It provides superior texture fidelity, complex structural adherence, and serves as the perfect foundation for fine-tuning or training custom LoRAs.

⚠️ IMPORTANT: This file contains ONLY the core diffusion weights (UNet/DiT). Text encoders and VAE are not included and must be loaded separately in ComfyUI:

1. Text Encoder (Required): Requires Qwen3-VL 4B (set type to 'krea2' in ComfyUI CLIPLoader)

2. VAE (Required): Requires qwen_image_vae.safetensors

Recommended Inference Settings:

* Steps: 28 - 34 (Standard diffusion trajectory)

* CFG / Guidance: 3.5 - 5.5

* Sampler: euler / er_sde

* Scheduler: simple or beta

* Resolution: 1024px – 2048px

Prompting Guidelines:

* Prompting Style: Write in full, descriptive sentences (Natural Language). Avoid comma-separated tags or keyword lists.

* Negative Prompt: Optional but can be used for style filtering.

Derived from Krea 2 developed by Krea.ai. Subject to the Krea 2 Community License Agreement.

Recommended Upscaling Settings:

* Upscale Model: 4x-UltraSharp (highly recommended for RAW photo fidelity) or 4x_NMKD-Superscale (for digital art)

* Upscale Method: Ultimate SD Upscale (Tile-based)

* Denoise Strength: 0.30 - 0.35 (Strict threshold to maintain original composition)

* Upscale Steps: 16 - 20

* Upscale Sampler/Scheduler: euler / simple