houseofboern/upscale-klein9b-edit-lora

Upscale / Detail edit LoRA — FLUX.2 Klein 9B (v2, epoch 7)

An edit LoRA: give it a soft or compressed photo as the reference image, and it re-renders the same picture with more detail.

• File: upscale_klein9b_edit_v2_e7.safetensors (rank 16, about 83 MB)
• Trained on: FLUX.2 Klein 9B Base, using 140 before/after pairs (soft original → detail-enhanced), with Fizgig's "Old Reliable" preset (rank 16, lr 1e-4) at 1024px
• Previews were rendered on the distilled FLUX.2 Klein 9B
• How to prompt it: see prompt.md

Honest notes

• Training seemed to stagnate. There is a clear change from epoch 1 to epoch 7, but very little from epoch 5 to 7. The run was stopped at epoch 7 of 55 for that reason. The loss stayed flat at about 0.53–0.56.
• It boosts colour saturation. The "after" images in the dataset were made by ChatGPT, which added saturation, and the LoRA picked that up. Expect richer colour, and lower the LoRA strength or correct the colour afterwards if it's too much.
• It gives the biggest gain on genuinely soft or low-detail inputs. On images that are already sharp, the change is modest.