ZIT-GGUF-dAIver-v1.5

Optimized Low-VRAM Workflow for Z-Image-Turbo (GGUF) with CacheDiT Acceleration Roughly based on WikkedAI’s WikkedZITv4 – refined and enhanced by Experimental_dAIver This workflow delivers the full power of Z-Image-Turbo in GGUF format, specially optimized for GPUs with less than 8 GB VRAM, like my RTX 4050 with only 6 GB. The integrated CacheDiT_Model_Optimizer and SageAttention2 (both optional!) provides a noticeable 1.4–1.6× speed boost with almost no quality loss. Two intelligent upscaling stages, automatic trigger-word integration via the Super LoRA Loader, and an extended save node complete this elegant setup. Version 1.5 brings significant improvements in speed, usability, and upscaling quality — while remaining extremely VRAM-efficient (tested on RTX 4050 with only 6 GB). What’s new in v1.5: PatchSageAttentionKJ integration for automatic Sage Attention optimization and faster sampling on supported hardware selectLatentSizePlus — intuitive aspect-ratio and resolution selector with beautiful presets (including 7:12 Tall Vista and other golden-ratio-friendly options) plus easy orientation swap Full SEEDVR2 Video Upscaler Subgraph — powerful DiT-based (.safetensors or GGUF) high-end upscaler that delivers stunning 4K+ results with intelligent resolution handling, Lab color correction, and temporal settings. Works exceptionally well on still images too, producing superior detail and coherence Main model updated to the higher-quality z-image-turbo-Q8_0.gguf CLIP switched to the abliterated Qwen3-4B-Instruct-2507.Q5_K_S.gguf (lumina2 type) Improved workflow organization, expanded notes, and more robust saving options Required Custom Nodes (updated for v1.5): ComfyUI-GGUF - https://github.com/city96/ComfyUI-GGUF - UnetLoaderGGUF + CLIPLoaderGGUF ComfyUI-CacheDiT - https://github.com/Jasonzzt/ComfyUI-CacheDiT - CacheDiT_Model_Optimizer – the turbo boost for DiT models nd-super-nodes - https://github.com/HenkDz/nd-super-nodes - NdSuperLoraLoader with tags, trigger words & beautiful UI save-image-extended-comfyui - https://github.com/thedyze/save-image-extended-comfyui - Advanced saving with metadata & dynamic filenames ComfyUi-MzMaXaM - https://github.com/MzMaXaM/ComfyUi-MzMaXaM - selectLatentSizePlus ComfyUI-SeedVR2_VideoUpscaler - https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler - SEEDVR2 Video Upscaler Subgraph Models & Downloads (exact paths) The following list explains the base models I am most frequently using with this workflow. The list as well explains where to put each file after you downloaded it. 1. Main Model (Diffusion Model) File: z_image_turbo-Q8_0.gguf (higher quality Q8_0) Download: https://huggingface.co/jayn7/Z-Image-Turbo-GGUF/resolve/main/z_image_turbo-Q8_0.gguf Target folder: ComfyUI/models/diffusion_models/ 2. Text Encoder (CLIP) File: Qwen3-4B-Instruct-2507-abliterated.Q5_K_S.gguf Download: https://huggingface.co/mradermacher/Huihui-Qwen3-4B-Instruct-2507-abliterated-GGUF/resolve/main/Huihui-Qwen3-4B-Instruct-2507-abliterated.Q5_K_S.gguf Target folder: ComfyUI/models/text_encoders/ (or clip/) 3. VAE File: ae.safetensors (~335 MB) Download: Usually included with Z-Image-Turbo setups or available here: https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors Target folder: ComfyUI/models/vae/ 4. Upscalers 4× Upscaler: 4xLSDIRplusN.pth (variant of 4x-UltraSharp) → https://civitai.com/models/116225/4x-ultrasharp 1× Skin-Contrast Upscaler: 1xSkinContrast-High-SuperUltraCompact.pth Download: https://huggingface.co/notkenski/upscalers/blob/main/1xSkinContrast-High-SuperUltraCompact.pth Target folder: ComfyUI/models/upscale_models/ 5. SEEDVR2 Models (for the new high-end upscaler – optional but recommended) : DiT Model: seedvr2_ema_3b-Q8_0.gguf VAE: ema_vae_fp16.safetensors Download from the official ComfyUI-SeedVR2_VideoUpscaler repository or Hugging Face and place in the folders required by the custom node. Key Nodes & Their Functions CacheDiT_Model_Optimizer + PathchSageAttentionKJ → next-generation turbo boost NdSuperLoraLoader with automatic trigger-word detection and clean tag interface selectLatentSizePlus → effortless aspect-ratio and resolution control KSamplerAdvanced with proven settings Two-stage classic upscaler (4× LSDIR + 1× Skin-Contrast) in its own subgraph, or New SEEDVR2 Video Upscaler Subgraph → for ultimate quality (optional, easily bypassed) SaveImageExtended with full metadata and dynamic filenames Recommended Settings (already set in the workflow) Sampler: res_multistep or dpmpp_sde Scheduler: beta (or ddim_uniform) Steps: 8–11 CFG Scale: 1.0–1.5 Shift: 4–7 Resolution: 864×1280 (portrait) – perfectly balanced for the golden ratio and typical ZIT outputs - the upscaler will automatically upscale by a factor of 4 How to Use the Workflow Install all required custom nodes Load the workflow Enter your positive prompt (NdSuperLoraLoader automatically adds trigger words) Adjust the negative prompt Choose your desired aspect ratio and resolution in the Size Selector Keep CacheDiT and Sage Attention enabled → Generate Optionally run the SEEDVR2 upscaler for breathtaking 4K+ results Done — images saved with complete metadata Special thanks to @LumaRift who provided the SeedVR2 subworkflow and some good advise on optimizing my setup.