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

1. Install all required custom nodes

2. Load the workflow

3. Enter your positive prompt (NdSuperLoraLoader automatically adds trigger words)

4. Adjust the negative prompt

5. Choose your desired aspect ratio and resolution in the Size Selector

6. Keep CacheDiT and Sage Attention enabled → Generate

7. Optionally run the SEEDVR2 upscaler for breathtaking 4K+ results

8. Done — images saved with complete metadata

Special thanks to @LumaRift who provided the SeedVR2 subworkflow and some good advise on optimizing my setup.