Training a Z-Image Turbo LoRA from 24 cat photos: full AI Toolkit settings and why 2,750 steps beat 3,000
meshive · reddit · 2026-11-05
The author trained a Z-Image Turbo LoRA on 24 photos of their cat on a single RTX 5090 and used it in ComfyUI, sharing a fully reproducible recipe:
- Dataset: 24 photos mixing poses, angles and backgrounds with clear facial markings
- Training: AI Toolkit, LoRA rank 32, 3,000 steps with checkpoints every 250, trigger word "idu", 1024x1024 previews (CFG 1, 9 steps)
- Generation: 1024x1024, 8 steps, CFG 1, resmultistep sampler + simple scheduler, model sampling shift 3.0, per-image LoRA strength (0.85 model / 1.0 CLIP for the 2,750-step comparison)
- Workflow: route the base model through the LoRA loader into the model sampling node, pass the text encoder through the loader too
- Surprise finding: the final 3,000-step checkpoint looked worse than 2,750 steps; caveat — comparison images used different seeds, so fair tests need fixed prompt, seed, sampler and strength
A 10-minute video walkthrough is included.
Related event: Training a Z-Image Turbo Cat LoRA with Just 24 Photos on a Single RTX 5090(2 posts)→
More from coding & agent
- Google donates Agent Substrate, gVisor and OpenRL to CNCF; Agent Substrate hits 1.0 on GKE — HowDevelop · 2026-11-06
- Deep-dive report on parallel coding agents: worktree isolation doubles costs, merge conflicts dominate — weswinder · 2026-11-06
- OpenAI's Justin Uberti to reveal GPT-Live speech-to-speech architecture at VapiCon 2026 — juberti · 2026-11-06
- Approved action ran twice in 3 agent frameworks: crash-resume test — KaleidoscopeNew9259 · 2026-11-06
- Before automating Shopify flows, run a dry run listing every product and price it would change — shashib · 2026-11-06
- Melite open-sources agentic ComfyUI: agents that generate and extend their own harness — data-cypher-000 · 2026-11-06