Krea2 LoRA training on ModelScope and testing in ComfyUI

This Russian tutorial trains a Krea2 LoRA on ModelScope, then tests it in ComfyUI and on the service's image generation page. Training runs on ModelScope's hosted resources rather than the viewer's computer. The registration walkthrough uses a phone number, SMS verification and a password, and shows an English interface option.

For the face LoRA example, the presenter uploads 10 images, sets a trigger word and accepts the recommended 6,000 training steps. In ComfyUI he loads and compares the 1,000-step and 6,000-step checkpoints. The later checkpoint produces a noisy background at weight 1, which he attributes to overtraining. Reducing the weight to 0.8 improves his example. He suggests trying 2,000 steps, but does not demonstrate that checkpoint.

Lowering LoRA weight also changes the balance between facial resemblance and a drawn style. He tests another LoRA trained elsewhere and recommends including drawn images in the dataset for that use case. Forge Neo compatibility is mentioned; no Forge Neo workflow is demonstrated.

ModelScope's generation page lets him select a checkpoint and test a prompt before downloading it. The presenter warns that the hosted training route he used makes the resulting LoRA public and downloadable by others. The hosted training and online testing are distinct from the local ComfyUI generation shown here.