Z-Image-Yaoguang Edition-Ultra-Realistic Detail Enhancement [A Chai]

Organized version: Section color coding (Red = User Input / Purple = Model Loading / Cyan = Condition Encoding / Blue = Processing / Green = Output), each section includes a sticky note with instructions, input nodes are marked in red with Chinese titles. Parameters and connections match the original; simply modify the red nodes to run.

· [Bilibili/YouTube: Work-Fisher] Author info section, for attribution display only, does not participate in the image generation process. No adjustable parameters, keep as is.

· [Work-FIsh-Z-Image-Yaoguang Edition (Thanks to KOOK master)] Workflow title and acknowledgement section, for description display only, does not participate in actual computation. No adjustable parameters, generally can be ignored.

· [Switch] Use the rgthree master switch to quickly enable/disable functional areas (such as switching SFW and NSFW models, toggling SeedVR upscale). Check the blocks you want to run as needed; does not involve generation parameters.

· [Z Sampler] Core process area: Positive/negative prompts are encoded and sent to KSampler for image generation, then decoded via VAE to save the original image. Adjustable sampling steps (default 8 steps, combined with acceleration LoRA, lowering further is not recommended), CFG (default 1), random seed (randomize for each face swap); positive and negative prompts can be modified as needed, remaining connections do not need to be changed.

· [1SeedVR Upscale] Uses SeedVR2 to upscale and repair details of the generated image, including loading the upscale model, scaling, upscale sampling, before/after comparison, and saving. Adjustable target pixels (default 6000), scale factor (default 0.3), seed; if VRAM is insufficient, lower the target resolution, and select model files as needed.

· [NSFW Model] Loads the NSFW base model (moodyPornMix_v10DPO) for adult image generation, switched on/off via the right-side switch. Generally, you only need to select the base model file here, keep the rest as default, no modification needed.

· [LoRA] Overlays two LoRA: Yaoguang style LoRA (Kook_Zimage_Yaoguang, default weight 1.0) responsible for the art style, and acceleration distillation LoRA (8-step Distill, default 0.45) responsible for fast generation with few steps. To enhance the style, fine-tune the Yaoguang weight; acceleration LoRA should generally be kept at default.

· [1Simulated Grain Texture] Overlays subtle film grain on the generated image to reduce AI plastic feel and increase realistic texture. Adjustable grain strength (default 0.015), grain saturation (default 0.45), scaling scale (default 4); increase strength slightly for a rougher look, generally keep at default.

· [1Adjust Image Contrast and Saturation] Post-processing color grading for the generated image, darkening highlights, increasing contrast, and lowering saturation to make it look like a casual smartphone snapshot. Adjustable brightness (default 0.9), contrast (default 1.2), saturation (default 0.8); fine-tune to personal preference, or use without changes.

· [SFW Model] Loads the regular Z-Image base model (z_image_bf16) for standard realistic image generation, controlled by the switch as one of two options with the NSFW model. Usually, you only select the base model file here, and keep the weight type as default.

· [Simple Text Input] Text box for entering image quality and style prompts; contents are sent to the positive encoder. Feel free to modify these prompts (default emphasizes ultra-clear realism, casual smartphone snapshot, voice-activated lighting atmosphere); this is the most frequently changed part for each generation.

· [CLIP Model] Loads the text encoder CLIP (qwen_3_4b, type lumina2), responsible for converting prompts into conditions the model can understand. Simply select the corresponding CLIP file, keep type and device as default, generally no changes needed.

· [VAE] Loads VAE (z_image_turbo-vae), responsible for converting between latent and images, determining color details of the generated image. Just select the VAE file matching the base model, no parameter adjustment needed, generally no changes required.

· [Resolution] Sets the canvas size and batch count for image generation; uses portrait 1080×1920, 1 image per batch. You can change width, height (default 1080×1920), and batch count; adjust for landscape or other ratios here—note that higher resolutions consume more VRAM.

New users register to get 500 RH Coins: https://www.RunningHub.ai?inviteCode=rh-v1302

This content contains AI-generated imagery and audio.