👉 Updated to v4.0 on February 24, 2026 (see change log at end of Description). Thanks to all for your helpful feedback!
Transform your images with a powerful Z-Image Turbo checkpoint and also upscale them — for relatively cheap! Typically, upscaling a 1024x1024 source image by a factor of 1.25 costs only 1.5-3.5 credits/tokens 🤙
1️⃣ Upload your image.
For best results, use a 1MP source image (e.g., 1024x1024, 768x1152, 768x1344, 832x1216, etc.) or slightly larger. Large source images may cause errors if you try to upscale too high.
General rule: the larger the source image, the lower you should make the Latent Upscale factor. And of course the larger the upscale, the higher the cost 🤑
If the generated image isn't photorealistic enough, try increasing the denoise and/or adding photography-related keywords like film grain or other photography-related terms at the start of the prompt (see PROMPT TIPS below).
2️⃣ Enter the Main Prompt in natural, grammatically correct language. Z-Image prefers natural language to the CLIP safetensors (tag-style prompts) used with SD models.
PROMPT TIPS:
Want to recycle an old SD prompt? Copy/paste it, then click Enhance.
Want to generate something similar to your existing image? Click Abstract, tick the Natural Language checkbox, then upload your image.
For extra photorealism, add the name of a famous photographer + mm lens + film stock. For example: The photograph was shot by Herb Ritts on Kodachrome film using a 35mm lens. According to some, adding natural lighting also increases photorealism.
👉 NOTE: There is no Negative Prompt field, because Z-Index Turbo doesn't use negative prompts.
3️⃣ Select the Denoise: Choose a range from 0.15 (very similar to original image) to 0.30 (similar) to 0.5 (noticeable change) to 0.6 (significant change) to 0.70 (radical change).
4️⃣ Select the Sampling Shift: Think of the shift as a slider that determines how the model balances composition versus detail.
1.0 - Low Shift: The model focuses more on the overall structure and big shapes.
1.73 - Creator Default Shift: This is the "sweet spot" discovered by the AuraFlow creators (fal.ai) for the best balance of quality and prompt adherence.
2.0, 3.0, 4.0 - Structural: Strong, clean composition; very "safe."
2.5, 3.5, 4.5 - Detail-Biased: Adding texture while keeping the grid clean.
For human skin, I personally find a shift of 3.50 gives the most photorealistic results, but you may have your own preferences.
5️⃣ Select the remaining parameters and click GO 🚀
A word about seeds:
Manually type or paste the seed of a previously generated image to refine it
Click - or + to increase or decrease the seed number (the slightest change will create an entirely new image from scratch)
Click 🎲 to let the AI choose a random seed for the image
Non-PRO (Standard) users: Depending on the time of day, generation times can sometimes be long (15 minutes to several hours). Sometimes queuing up/stacking 2 generations in a row can help speed things up — plus the 2nd gen usually costs less because the tool is 'warmed up' 😉
Enjoy 😊
— ℝ𝕖𝕩𝕠
P.S. When you click on a rendered image, TA's metadata shows the checkpoint's ID instead of its name, so here's a cheat-sheet:
ID ending in *2838 = ZIT bf16
*8499 = ZIT fp8
*8233 = ZIT REFINE fp16 v0.2
*1987 = ZI Ultra - Turbo v1
*8685 = Z-Imagibility Turbo v1.0
*9875 = Z-Image Cinema soft v2.0
*2624 = Z-Image-Turbo-Enhanced v1
*1415 = z - full - image (by Stable Yogi - "Lab Rat v0")
*9781 = LexiVision-Z
*8760 = Diving-ZImageTurbo-Real Asian v5
*5076 = Dark Beast Z | 黑兽⚡️DBZ vDarkBliZ6
*2813 = REAL PHOTOGRAPHY
*1300 = Z-Image Turbo SFW [Photorealistic] v2.1_BF16_AIO
*4367 = Nova Reality ZI - v1.5 Turbo full bf16
*9830 = Z-image-Turbo-Pro-bf16 v6
*4583 = ₳M₳TR - TA_Excl-V2.0
*4714 = ZIT Asian Utopian
*2944 = unStable Revolution ZIT v2.0
*8374 = Jib Mix ZIT - v2.0
*2266 = ZIT Uncnsrd v3
*7874 = rRealisMAX Z-Image ZIT v0.1
*8004 = Z-Image Pro - v1
*2711 = BEYOND REALITY bf16
*9084 = CyberRealistic ZIT v0.5
*************
Change log:
v4.0 - 24 Feb 2026 - Removed number of steps 5, 6, 7, and >12, as results were uneven; removed Latent upscale sizes 1.05, 1.33 and 1.66 for the same reason; updated and expanded number of available checkpoints based on popularity on TA & Civitai. Thanks again for the helpful feedback 🙏
v3.0 - 21 Jan 2026 - Corrected a misconfiguration (my bad!) which had been compressing some uploaded source images; added more Steps options; optimized Sampling Shift and Denoise.
v2.0 - 11 Jan 2026 - Added Model Sampling options; optimized latest upscale quotients; changed sampler/scheduler to res_multistep + sgm_uniform. Started work on a separate variant of the tool which will instead use pixel upscale (SeedVR2).
v1.0 - 2 Jan 2026 - Initial release