2603 ZIT — By Stable Yogi
Fast photoreal Z-Image Turbo with a natural, real-life look — believable skin, faces, and lighting instead of the over-sharp "stock studio" feel. ~8-step renders, SFW → spicy.
Highlights
📷 Natural, real-photo realism — soft, believable lighting
🧬 Clean skin texture, crisp faces, reliable anatomy (holds up on complex poses)
🎨 Versatile — portraits, full-body, editorial, candid
🧩 Excellent LoRA base
What's new
Retuned for true-to-life realism — dialed back the over-processed studio look
Ships in fp8 (Forge + ComfyUI) and Q8_0 / Q4_0 GGUF (ComfyUI)
Recommended settings
Steps 8–9 · CFG 1.0 · Sampler Euler · Scheduler Simple
Resolution 896×1152 or ~1024 square
Which file — Q8_0 GGUF = best quality (ComfyUI) · fp8 = Forge or ComfyUI · Q4_0 = smallest / low-VRAM
Want the stronger, higher-detail Pro build? → Get the Pro version of my models here
Description
2603 ZIT — By Stable Yogi
Fast photoreal Z-Image Turbo with a natural, real-life look — believable skin, faces, and lighting instead of the over-sharp "stock studio" feel. ~8-step renders, SFW → spicy.
Highlights
📷 Natural, real-photo realism — soft, believable lighting
🧬 Clean skin texture, crisp faces, reliable anatomy (holds up on complex poses)
🎨 Versatile — portraits, full-body, editorial, candid
🧩 Excellent LoRA base
What's new
Retuned for true-to-life realism — dialed back the over-processed studio look
Ships in INT8
To run this INT8 on forge you will need this forge extension
Recommended settings
Steps 8–9 · CFG 1.0 · Sampler Euler · Scheduler Simple
Resolution 896×1152 or ~1024 square
Which file — Q8_0 GGUF = best quality (ComfyUI) · fp8 = Forge or ComfyUI · Q4_0 = smallest / low-VRAM
Want the stronger, higher-detail Pro build? → Get the Pro version of my models here
FAQ
Comments (12)
wow this new 2603 fp8 version is amazing! i get incredible results! Thanks again Stable_Yogi! great work!
<3
Appreciate that! The fp8 was the version I really wanted to get right for full-quality setups, so glad it's delivering for you 🙏
Easily one of the best zit model. Possible to also get a bf16 variant for 2603?
No BF16 , but INT8 coming soon.
Hey Yogi! I just ran a full stress-test experiment with your Qwen3-4B-Q4_K_M.gguf text encoder on my RTX 5050 8GB setup (paired with Z-Image-Turbo GGUF), and I wanted to share the results immediately.
I built a complete, interactive visual novel web interface running on a local Python backend. The prompt processing is fully modular: it can either run a local LLM (Gemma-4 via Ollama) or leverage Google's Gemini API to expand prompts, then pushes the rich English text via Python scripts straight to the ComfyUI API.
In my chat UI, I implemented a strict hardware-saving logic with two separate buttons that perfectly leverage your model:
1. "Left Button" (Face Avatar 384x512) - The script automatically disables the hands-fix LoRA to save resources. Hot-start generation takes a stable 15.4 seconds.
2. "Right Button" (Story Scene 768x1024) - This is where your GGUF text encoder saved the day at a higher resolution.
Here is exactly what your Q4_K_M GGUF quant changed compared to the heavy FP8 version at 768x1024:
• VRAM Saved: It freed up a solid 1.0 - 1.4 GB of VRAM during generation (dropping from a dangerous 6.7 GB ceiling down to a comfortable 5.8 GB).
• Thermals Drop: Peak GPU temperature dropped by 5°C (from a thermal-throttling 82°C down to a safe 77°C), and fan speeds decreased noticeably.
• Blazing Speed: The "hot start" generation time settled at 25 seconds for the high-res story scenes. ComfyUI keeps the un-quantized layers resident in RAM, meaning zero speed penalty on repeated generations.
• Quality: Visually, there is 0% loss in final image details, skin textures, or character anatomy.
Your quant became the perfect engine for my interactive text game, completely saving my 8GB card from Out-of-Memory crashes on heavy story scenes. Amazing work, thanks a lot!
My project web interface screenshot:
Wow... this is one of those comments that makes all the late nights worth it. 😄
First off, thank you for taking the time to actually stress-test it instead of just giving it a quick spin. The VRAM, thermals, and hot-start numbers are exactly the kind of real-world data that's incredibly useful.
I also love the way you've built your pipeline. Splitting the avatar and story scene generation into separate resource profiles is a really smart approach for 8GB GPUs. That's the kind of optimization that makes AI tools actually enjoyable to use instead of constantly fighting OOM errors.
And hearing that the Q4_K_M encoder shaved off around 1–1.4GB of VRAM while keeping image quality intact is exactly what I was hoping these quants would achieve.
Seriously, thanks for sharing all the details—and the screenshot too. It's always awesome seeing people build completely different projects around these models. Wishing you the best with the visual novel; it sounds like a really cool project!
Stable Yogi keeps his touch from one checkpoint family to the other. Stay awesome brother, you're a pillar of the community.
Means a lot, brother — thank you 🙏 I just try to keep each release honest and useful, and the community's the reason it's worth doing. Glad to have you around.
Stable Yogi is an absolute unappreciated genius! This INT model is pure black magic and a technological masterpiece.
I did a head-to-head test on my new RTX 5050 setup, and the results are mind-blowing:
Standard run: took 272 seconds (12.42s/it).
INT version (using the native convrot mixed precision): cut the time down to just 132 seconds (6.29s/it)! It literally cut the generation time in half!
VRAM management is flawless, leaving plenty of room for heavy workflows. Note to everyone: since this card runs on the brand-new Blackwell architecture, you absolutely need to manually update your PyTorch to the new CUDA 13.0 nightly environment (which dropped just a few days ago) to unlock these blazing fast speeds.
People are seriously sleeping on this model. Massive respect to the author for this incredible optimization! 🔥
Ah man, thank you, this genuinely made my day. Stoked it's flying on your setup like that. I keep the Pro versions over on my portal, if you ever wanna dig deeper. Appreciate you 🙏
Here is the custom node component for your model! 🤝
I have packed everything into a single, clean component for ComfyUI v3. It combines Checkpoint, GGUF CLIP, and VAE Loader into one compact block to keep your workflow clean and wireless.
How to use:
Copy the JSON code below.
Create a new text file on your PC, paste the code, and save it as GGUF_Super_Loader.json (make sure the file extension is .json, not .txt).
Move this file into your ComfyUI components folder: ComfyUI/user/default/components/.
Refresh your ComfyUI page. Now you can find this super-node by double-clicking on the canvas and searching for GGUF_Super_Loader.
Node Workflow Code:
json
{ "69:0": { "inputs": { "vae_name": "ae.safetensors" }, "class_type": "VAELoader", "_meta": { "title": "Загрузить VAE" } }, "69:1": { "inputs": { "ckpt_name": "zimageTurboByStable_2603INT8Convrot.safetensors" }, "class_type": "CheckpointLoaderSimple", "_meta": { "title": "Загрузить сheckpoint" } }, "69:2": { "inputs": { "clip_name": "Qwen3-4B-Q4_K_M.gguf", "type": "qwen_image" }, "class_type": "CLIPLoaderGGUF", "_meta": { "title": "CLIPLoader (GGUF)" } } }