Ideogram 4 LoRA training: Anime style in 3.5 hours on a 4090
Only 40 images, 3000 steps, 6-bit quantization—results are promising.
A developer known as kingroka is testing LoRA (Low-Rank Adaptation) training for Ideogram 4, the latest AI image generation model. Their goal was to overcome Ideogram's default tendency to render anime-style images with an oil painting aesthetic. By training a dedicated anime LoRA on a small dataset of only 40 images with JSON captions, they achieved promising early results in just 3,500 steps.
The training was performed on an RTX 4090 using the AIToolkit platform, taking approximately 3.5 hours. Due to memory constraints, the developer reduced quantization to 6-bit to avoid out-of-memory errors. The LoRA is not yet ready for release—the plan is to increase dataset quality and size, and double the training steps. Example outputs (not in the training set) show improved anime style consistency, with prompts like 'Mask Lady,' 'Noodles,' and 'Laptop Lady' producing more faithful anime aesthetics.
- LoRA trained for Ideogram 4 with only 40 images and 3000 steps.
- Training completed in 3.5 hours on an RTX 4090 using 6-bit quantization.
- Results show improved anime style over Ideogram's default oil-painting look.
Why It Matters
Efficient LoRA fine-tuning on Ideogram 4 opens custom style control for small-batch creators.