MOSS Image Sharpness Slider

Model description

MOSS-Image Sharpness Slider

A Concept Slider LoRA for MOSS-Image that moves an image between sharp and blurry by LoRA strength alone. Positive strength sharpens, negative strength blurs. No trigger word, no prompt changes.

IMPORTANT: Intended usage is deblurring the MOSS Base checkpoint. It will have a weaker effect on MOSS SFT, and almost none on SFT+Turbo. The reason is that SFT is already tuned to filter out the blur, and Turbo distills it further.

Usage

Load as a regular LoRA on MOSS-Image Base or SFT. Start at 1.0 for sharper output and -1.0 for blur, then adjust. Strength 0 is the untouched model. No trigger word.

Training

Trained with the `concept_slider` process in Ostris AI Toolkit on MOSS-Image Base (int8 convrot). The slider direction comes from the prompt pair; the dataset only supplies latents at both ends of the range.

• Dataset: 72 photos and the same 72 photos with Gaussian blur, 1024 px, uncaptioned. Default captions "sharp, in focus" and "blurry, out of focus".
• Positive prompt: "a sharp photo, perfectly in focus, crisp fine detail, clear edges"
• Negative prompt: "a blurry photo, blurred, out of focus, soft and smeared"
• Target class and anchor: empty. Guidance strength 3.
• LoRA rank 4, alpha 4. Layer 22 excluded from the LoRA; it is the skip-layer guidance block for MOSS, and `uncond_skip_layers: [22]` during training.
• lr 0.001, adamw8bit, weight decay 0.001, EMA 0.99, batch 1, flowmatch with shift timesteps, bf16.
• This is the step 400 checkpoint. Steps 350 and below had not converged; 400 gives the full effect without visible overfitting.

Notes

Each slider step runs several passes of the base model, so training is about 3x slower per image than a regular MOSS LoRA. 400 steps took well under two hours on a single consumer GPU.

Download model

Download them in the Files & versions tab.