Training an SDXL LoRA from SD1.5-generated images keeps failing — author asks for help
mfowler2442 · reddit · 2026-09-23
After nailing a character LoRA on SD1.5, the author tried migrating the character to SDXL using 30 upscaled SD1.5 images. Three LoRA attempts with varied Kohya settings captured the general resemblance but never the exact identity. Next steps: add more close-up headshots and reroll weak off-angle faces suspected of hurting training.
More from Multimodal
- Adobe completes Topaz Labs acquisition, bringing Emmy-winning AI enhancement to Firefly and Photoshop — rufusd · 2026-09-23
- Recraft launches V4.1 Flash, generating images in 1.3 seconds as fastest image model — aziz4ai · 2026-09-23
- Claude Opus 3D castle test shows dramatic quality jump in 7 months — chaseleantj · 2026-09-23
- invideo launches Agentic Sound Design: AI agents auto-score, mix and master your video audio — sachinmaya1980 · 2026-09-23
- 3D Gaussian splatting turns phone shots into explorable photorealistic 3D scenes — import_jmr · 2026-09-23
- New ComfyUI nodes let you mask any video area and insert characters — roychodraws · 2026-09-23