PickKromoss: Generate 1024px Images in 8 Steps with MOSS-Image Fine-Tuning

A text-to-image model fully fine-tuned from MOSS-Image-Base at 1024px resolution. It shares the same architecture, text encoder (Qwen3.5-9B), and VAE as the existing MOSS-Image-SFT, and has an identical repository layout, allowing it to be directly substituted into existing pipelines. It provides EMA weights trained on 7.04 million training images.

1 / 4

In addition to the default 50-step inference, two 8-step turbo LoRAs are distributed alongside it. The 8-steps-dmd2-lora, further trained with a DMD2 GAN stage, offers superior contrast and detail, recording a CLIP score of 28.13 on the PartiPrompts benchmark. Applying this LoRA enables fast image generation with CFG 1 and 8 steps.

It supports native ComfyUI integration and is available in two formats: a single bf16 file and an int8 quantized file. The int8 version has a size of 5.4GB, improving memory efficiency, and works immediately by simply replacing the file in existing MOSS-Image workflows. It is released under the Apache 2.0 license, allowing free use in commercial environments.