Over the past few days, I’ve been experimenting with on-demand GPU services for AI image generation and LoRA training, and the results have been impressive.
Initial Training Tests
I started by training Z Image Turbo on AI Toolkit using a rented RTX 6000, but I struggled to fully utilize the available VRAM. My initial approach—90 images, 3000 steps, batch size 1, 1024×1024 resolution—proved inefficient. Attempting larger batches with fewer steps didn’t yield better results, suggesting this setup is better suited for large-scale datasets (e.g., base models or checkpoints).
Switching to a 5090 GPU with batch size 2 and 1000 steps was far more effective—training completed in under an hour with much better stability.
ComfyUI Workflow & Multi-Pass Generation
After some initial setup challenges (model loading, workflow configuration), I got ComfyUI running smoothly on the 5090. Z Image Turbo alone generates a 1024×1024 image in just ~2 seconds—an incredible speed boost.
To refine output quality, I implemented a three-pass KSampler workflow:
Z Image Turbo (base generation)
SDXL Illustrious (ILL) (style application)
SDXL ILL Upscaling (final 2K resolution)
This pipeline produces high-quality 2K images in ~60 seconds, with Z Image Turbo excelling at prompt adherence—especially for camera angles, composition, and framing—while SDXL ILL applies my preferred semi-realistic anime aesthetic.
Automated Dataset Generation
To maximize diversity, I integrated wildcard lists (Character, Clothing, Environment, Composition) that dynamically generate randomized prompts. The workflow automatically:
Saves the wildcard-generated prompt as a .txt file
Exports the final upscaled image
Ensures caption-image alignment for easy dataset organization
This method is exceptionally efficient for building tailored datasets, especially for digital art styles. While its effectiveness for photorealistic models remains untested, the speed and quality for anime-style generation are game-changing.
With powerful rented hardware streamlining the process, this is now my go-to pipeline for scalable, high-quality dataset creation.
So for my next LoRA training run I want to include more diverse images in the dataset and I'll be excited to see the results later!