Chose high-quality images.They must be clear, sharp, not over saturated and at least 1024x1024. Crop and resize image to 1024x1024 if it too shrink
Number of images for quality lora is about 40-50 images with different angles and emotions, use different states of characters like sitting or running, smiling and so on.
Add description for every image. You can use QWEN VL, but always double-check what it writing about your image.
For example, ,my description for image.
A cheerful yellow fox with white tipped ears, a bushy white chest fur, and bright cyan eyes looks slightly to the side while extending his right hand outward in an open-palmed gesture as if explaining or welcoming someone. He wears brown leather straps across his torso secured by a brass buckle adorned with a crescent emblem, along with matching wristbands on both arms. His head features goggles pushed up onto his forehead revealing one orange lensed eye and he has two long tails curled upwards behind him. The character’s face displays a friendly, animated expression with wide-eyed curiosity and a slight smile hinting at playful confidence. Soft, even directional light illuminates him from above and front-left, casting gentle shadow contours beneath his chin, hands, and belt without harsh contrast. Shot from a medium close-up perspective near eye level, the framing centers the fox prominently, emphasizing his expressive posture and gear-heavy attire for maximum visual impact.
As you can see, QWEN VL made mistake -- His head features goggles pushed up onto his forehead revealing one orange lensed eye and he has two long tails curled upwards behind him. So I changed to His head features goggles pushed up onto his forehead revealing googles with red lenses and he has two long tails curled upwards behind him.
Double check description again. Delete words if there are too many of them.Make the description concise and short, but detailed enough.
Open (or install and open) Onetrainer. You can use onetrainer preset for qwen image with your vram, but I prefer to customize some options.
Add concept and select path
9. My settings for training Lora and some advises.
I use prodigy with cosine because it don't need learning rate. Just use 1.0. And it can produces decent quality. Always use cpu offloaded if it says on errorlog CUDA OUT OF MEMORY. It will take longer time but at least it can be trained. Use layer offload fraction from 0.1 to 0.9 until the error disappears. Layer filter attn-mlp, although with full or custom makes no difference. You can lower epoch count to increase speed and avoid over fitting. User 80-60, if it works for you.
Do not forget to sample every 8-10 epoch to make sure that the process is going in the right direction
Backup every 8-10 epoch in backup tab to not over fitting.
Lora settings tab. Lora rank and Alpha for qwen image can be more lower like 8/16. Iam using dropout prop at 0.1.
Start training. I am not quantize layers and chose the most precise bf16 but if you have low VRAM you will have no choice,
At the end in comfyui with dpm adaptive normal you can generate something like this
You can change model strength to 0.7 or even to 0.37 to change appearance
But the most important thing remember -- you can always change some of your mistakes in QWEN IMAGE EDIT OR FLUX IMAGE EDIT like this
My presets in attachments. Feel free to comment. Sorry for my bad English.
Training lora for Qwen Image as example in onetrainer. My experience.
0
0