MOSS-Image Sharpness Slider LoRA: what -1 and +1 do to our five standard pictures

Worked2026-10-29 · MOSS-Image SFT, MOSS-Sharpness-Slider · RTX 5070 12GB

A tiny LoRA (18 MB) for MOSS-Image went around on X this week: no trigger word, no prompt, only the strength changes the picture. Positive strength sharpens, negative strength blurs. We tried it on our standard five prompts, because "one slider, no prompt" is easy to test fairly. Thanks to @aiaicreate for the pointer; the LoRA is aoleg/MOSS-Sharpness-Slider (Apache-2.0 on the model card).

The test

MOSS-Image SFT (int8, 25 steps, er_sde, cfg 2.5) with the LoRA at strength -1 and +1, the five standard prompts (portrait, landscape, anime, text, busy night market), seeds 300 and 301, 1024×1024, one rented RTX 5070 12GB. The "no LoRA" column is our existing SFT run with the same settings.

What we saw (judged by eye, seed 300)

• The slider does what it says. At -1 the backgrounds and fine texture are softer (the bakery behind the OPEN sign is clearly out of focus, the landscape loses some crispness). At +1 skin pores, fabric and line work look crisper.
• It also changes the picture. With the same seed, the composition moved in every row: a different pose and camera angle for the anime girl, a bigger sign in the text scene, a different crowd in the market. So this is not a pure "sharpen this picture" filter; compare styles, not pixels.
• Text stayed correct. "OPEN" is spelled right in all three columns.
• Faces: the old fisherman looks about as natural at both strengths; at +1 his skin looks rougher.

How much to trust this

• One seed viewed per cell, no numeric sharpness measurement, one machine.
• The LoRA is described as aimed at removing blur from the Base model; we only tried SFT.
• We did not test intermediate strengths.