Training a Z-Image Turbo LoRA from 24 photos: why 2,750 steps beat 3,000

meshive · reddit · 2026-11-05

The author trained a Z-Image Turbo LoRA on 24 cat photos using AI Toolkit on a single RTX 5090, then used it in ComfyUI, with a 10-minute video walkthrough (prompts in the description).

Training: dataset mixed poses, angles and backgrounds showing face/markings clearly; LoRA rank 32; 3,000 steps with a checkpoint every 250; trigger word idu; 1024x1024 preview samples every 250 steps at CFG 1, 9 sampling steps.

Generation (ComfyUI): 1024x1024, 8 steps, CFG 1; resmultistep sampler, simple scheduler, denoise 1.0, model sampling shift 3.0; LoRA strength tuned per image (model 0.85 / CLIP 1.0 for the 2,750-step comparison). Route the base model through the LoRA loader into the model sampling node and pass the text encoder through the loader too.

Key finding: the final 3,000-step checkpoint looked worse than 2,750 — overshot. Caveat: comparison images used different seeds, so part of the difference is randomness; fair tests need fixed prompt, seed, sampler and strength. Everything is open source / open-weight, so you can replicate on your own GPU. The author discloses running Meshive, the GPU cloud used, but says the settings are platform-agnostic.

Original post →

Training a Z-Image Turbo LoRA from 24 photos: why 2,750 steps beat 3,000

More from Multimodal