ALWAYS LOOK INTO MY SAMPLES BEFORE DOWNLOADING TO UNDERSTAND IF THAT VERSION MEET YOUR EXPECTATIONS. Recommended generation parameters can be easily extracted in ComfyUI by dropping the sample into its window.This is the CinEro NG Krea 2 series of checkpoints. Aimed to deliver the dark cinematic SFW / NSFW picture with unrestricted Qwen 3 VL Heretic Text Encoder. Trained with sci-fi horror themed moody portraits dataset. Developed mostly to be used in ComfyUI, but should work in other Stable Diffusion generation software.<h2 id="versions-info">Versions info</h2><h3 id="v2g-turbo">v2g Turbo</h3>Recommended settings for Text-to-Image:<ul><li>steps 7..9;</li><li>sampler "exp_heun_2_x0_sde" + "simple";</li><li>CFG 1.0..1.1;</li><li>initial resolution 0.75..1.6 MPx.</li></ul><h3 id="v2b-turbo">v2b Turbo</h3>Recommended settings for Text-to-Image:<ul><li>steps 7..12;</li><li>sampler "exp_heun_2_x0_sde" + "simple";</li><li>CFG 1.0..1.3;</li><li>initial resolution 0.75..1.2 MPx.</li></ul>Euler and ER_SDE samplers give grid pattern. If you use ComfyUI I recommend adding the "ModelSamplingAuraFlow" node with "shift" parameter in 0.8..1.7. It probably also helps managing grid pattern.<h3 id="v2a">v2a</h3>BF16 is contrast and sharp, has more details in background. FP8 made by quantizing all tensors into FP8. My quick tests showed that FP8 sometimes give a textures with barely noticeable regular grid pattern (repeating grain noise on skin or sky, for example). It is needed to focus on textures to see, but sometimes noticeable. Let me know if you see such artifacts or other issues.Recommended parameters and workflow can be extracted from my samples.👇 READ THIS PLEASE 👇I prefer 2-stage rendering. Text-to-Image at 0.75..1.2 MPx resolution (the lower resolution, the less stuff in scene) with any compatible sampler (Euler, sa_solver), er_sde, dpmpp_2m, seeds_2, seeds_3, gradient_estimation, exp_heun_2_x0_sde). Second pass (HiRes Fix) I'm doing with Details Daemon node. Krea 2 Turbo have a very tight CFG range (0.9..1.5). You cannot adjust contrast and details using CFG knob, because you need to be much more precise. Details Daemon let you smoothly regulate CFG scale depending on Time Step.<h3 id="v1g-int8-cobvrot">v1g INT8 COBVROT</h3>Turbo workflow (10 steps) ==&gt;&gt; <a target="_blank" rel="ugc" href="https://civarchive.com/images/141569765">https://civarchive.com/images/141569765</a>Made from v1e (All in One) by applying more fine-tuning and quantization script to convert it into INT8 CONVROT. For me it looks less contrast (better use 20 steps and CFG 3..5 or more). Also, it looks less stable. But twice smaller file. Let me know please if that format works for you or you have any problem.performance: someone might expect the speed increase, but unfortunately speed on 4060 TI and 5060 TI looks the same; below is the DEBUG output that lists the exact grouping of all tensors that depends on data bits of weight encoding. <span style="background-color:rgb(26, 27, 30);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:16px">[SaveAsSafeTensor] DEBUG: Tensors: 879, Dtypes: {'torch.bfloat16': 166, 'torch.float16': 40, 'torch.int8': 224, 'torch.float32': 225, 'torch.uint8': 224}<span style="background-color:rgb(26, 27, 30);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:16px">Roughly the half of the weights (insensitive ones) were encoded in INT8. Other tensors must be slower as they were encoded in FP32 or BF16. troubles:<ul><li>At least one user have problems with this version in ForgeNeo (don't know which version of it was used); looks like ForgeNeo do not properly recognize which loader needs to be used for Krea2 Turbo INT8 Convrot.</li></ul><h3 id="v1e-all-in-one">v1e All in One</h3>It also has Qwen3VL Heretic Text Encoder and VAE baked in. Just use my workflow or use a regular Load Checkpoint node in ComfyUI combined with a regular KSampler. Nothing tricky needed. If you adapt the SDXL HiRes Fix technique described below, you may get better textures, but with v1e even single Text-to-Image pass at 1MPx resolution works well for me. Hope you will get good results also in your setup.Recommended settings for Text-to-Image pass:1216x832 (or 832x1216), Exp_heun_2_x0_SDE sampler (Simple, Normal, Beta, KL_Optimal schedulers), Steps 12, CFG 1.0...1.2.<h3 id="v1">v1</h3>It has Qwen3VL Heretic Text Encoder baked in. So, by using this model you take full responsibility on the resulting safety of the rendered images. Qwen Image VAE also embedded. CLIP and Unet parts both quantized to FP8.

CinEro-NG-Krea2

Little bit of additional fine-tuning.INT8 CONVROT, FP8, Q6 GGUF, NVFP4 - all tested on RTX 5060 Ti in ComfyUI. INT8 and FP8 tested on RTX 4060 Ti.

Files with SHA256: f8481457390e1813f442e0569ec28c370c7567597a414801aa4a7607d457edc5