MOSS Image

The 8-step LoRA has two versions, SFT and Base, and SFT should be chosen for everyday use; the 4-step LoRA should directly use the 2.0 version, avoiding the old 1.0 version.

A Bilibili uploader conducted a systematic comparative test of two acceleration LoRAs for MOSS Image (4-step and 8-step versions), using the native 25-step generation as a horizontal reference.

The test used a three-image comparison workflow, running the native 25-step, 4-step LoRA and 8-step LoRA versions simultaneously; operationally, one only needs to load the corresponding LoRA after the model, change CFG to 1, set the steps to 4 or 8, and use Euler simple uniformly as the sampler.

Overall test recommendations: choose the 4-step 2.0 version for speed; switch back to the native 25-step when ultimate detail is needed; 8-step suits specific aesthetic scenarios but watch out for composition risks.

In anime-style multi-character scenes, the native 25-step showed obvious errors in background areas, the 4-step version was more stable, but the 4-step generated images had a hazy feel and lacked detail; the 8-step showed background misalignment problems.

For scenarios such as tattoo images that need to preserve the artist's brushstrokes, the native 25-step preserves more of the artist's effect, while the 4-step weakens this stylistic feel; but for manga-type content, the impact of four steps is almost negligible.

MOSS Image has about 5B parameters, uses the Apache 2.0 license, is commercially usable, supports only text-to-image, and does not support natural-language editing.