Do you notice that variety collapses when training Style LoRAs on modern models like Qwen and Flux Klein? What's worked for you?
A user on Reddit is seeking advice regarding a specific issue encountered when training style LoRAs on newer image generation models like Qwen-Image and Flux Klein. The problem is a collapse in compositional variety, where generated images maintain similar layouts and subject positioning despite variations in color and detail. The user has experimented extensively with inference-side techniques and training configurations but has not found a definitive solution, particularly for flow-matching architectures that commit to composition early in the denoising process. They are looking for community insights on dataset structure, captioning strategies, or training configurations that could improve variety, and are also open to paid contract work for this production application. AI
IMPACT Users training custom models are encountering challenges with compositional variety, impacting the flexibility of generated outputs.