FFusion Core 2023
⚠️ Experimental — not a polished release.
Retrained a personal 2023 SDXL dataset (969 images + original captions) on Zimage Turbo using only Civitai Trainer — no local training, no manual caption editing.
🧪 Why This Exists
A stress test to see how Zimage Turbo handles heavy repetition — 20 epochs at just 2 repeats on a large abstract dataset. Deliberately overbaked. Results will vary wildly, and that's the point.
🎨 What It Does
Abstract backgrounds & gradient compositions
3D cube/container renders
Cloud & smoke formations
Colorful surreal portraits
Neon-lit geometric shapes
Best used as a style layer, background generator, or starting point for abstract designs.
⚙️ Training Config
Base Model Zimage Turbo
Engine ai-toolkit (Civitai Trainer)
LoRA rank 32 / alpha 32
Epochs 20
Repeats 2
Resolution 1024
Optimizer AdamW 8-bit
LR / Scheduler 1e-4 / constant
Text Encoder not trained
💡 How To Use
Trigger word: ffusion
WeightExpect 0.3–0.5
Subtle color/style bleed 0.5–0.8
Recommended range 0.6
Dataset and the original 2023 SDXL LoRA available on request.
Description
{
"lr": 0.0001,
"engine": "ai-toolkit",
"epochs": 20,
"ecosystem": "zimageturbo",
"keepTokens": 0,
"networkDim": 32,
"numRepeats": 2,
"resolution": 1024,
"lrScheduler": "constant",
"minSnrGamma": null,
"noiseOffset": null,
"networkAlpha": 32,
"optimizerType": "adamw8bit",
"shuffleTokens": false,
"textEncoderLr": null,
"flipAugmentation": false,
"trainTextEncoder": false
}