RAW Upscale MOSS

A detail and texture style LoRA for MOSS-Image, trained and tested with LoRA Dataset Studio.

Free download, available now. The training dataset on Patreon is a separate purchase.

One bee, seven strengths

Browse the gallery image by image to compare 0.0 / 0.7 / 1.0 / 1.4 / 1.8 / 2.5 / 3.5. The LoRA strength is printed directly on every image. All seven renders use the same prompt, seed and generation settings; 0.0 is the baseline with RAW Upscale MOSS disabled.

The series shows how increasing the weight changes fine texture, the bee's appearance, the flowers and the composition. Start around 0.7-1.0 for a moderate effect, then explore higher strengths for a stronger transformation. The examples are original LDS renders with only the strength label added.

Example setup: MOSS 4-step profile

• Base model: moss_image_sft_bf16.safetensors.
• Acceleration adapter: moss_image_sft_4step_v2.0.safetensors, strength 1.0, including in the baseline.
• The existing 4-step-profile renders record 6 sampling steps in their saved workflows.
• Euler / Simple, CFG 1.0, seed 1152340058, resolution 864 x 1536.
• Prompt prefix used: raw test upscale. The LoRA was trained in always-on style mode; the prefix is retained in the image metadata for reproducibility.

These are text-to-image examples. Upscale is the name of the detail-oriented style; output resolution is set by your workflow. The download contains this LoRA only. The MOSS base model and acceleration adapter are separate requirements for reproducing the gallery.

Training

OpenMOSS-Team/MOSS-Image-Base, 2,000 training steps, rank 16, alpha 16, 1024 training resolution. Final checkpoint in FP16 SafeTensors format.

Get the training dataset

For dataset purchases, visit my Patreon shop - LoRAperfectgf. Check the pack listing for availability, included files and terms. Dataset purchases are separate from the Civitai LoRA download.

Made with LoRA Dataset Studio. Please credit perfectgf when sharing your results.