Ultra-realistic model with cinematic flavor. Focus on realism, cinematic photography style(portraits and landscapes), moody lighting and atmosphere, anatomical features, expressions and body language, creativity (sci-fi, horror). Works very well as the refiner for Krea2 Turbo and Z-Image Turbo draft renders.

XL-V1

This version originated from my other CinEro XL models with the mixtures of FallenLeaf XL and XL-TCD-IOR models. Many iterations of deep training and additional merges performed on initial mixture.

Works best using two stage rendering. Text-to-Image with smooth sampler like Euler A or DDPM gives smooth / low noise composition. High Res Fix Image-to-Image pass with more contrast sampler (DPM++ 2S A, EXP_HEUN_2_X0_SDE, etc) renders the noisy / grainy textures and produces more contrast lighting.

BF16 variant gives almost the same composition and very similar quality / details as the FP32. FP8 differs significantly. It somtimes gives more cartoonish visuals, but I didn't test it extensively. Just expect FP8 to be noticeably different.

Special note on "Detail Daemon Sampler" in ComfyUI. This custom node probably can give you a chance to get details in single pass workflow (singly Text-to-Image pass) for OK-ish results. I tried this, but didn't tune it a lot to get the same quality as with T2I + HiResFix. Also, this version was trained in 1024 px resolution. So, I don't expect great performance in high resolution Text-to-Image (you may get duplicate fingers and other small objects when resolution is to high). Up to 2MPx resolution it may work OK, but no guarantee.