Faster image generation usually means accepting a quality trade-off, but Kromoss’s reported PartiPrompts scores suggest its 8-step route can do more than preserve quality: the DMD2 LoRA scores 28.13, above the full model’s 27.90, while using CFG 1 instead of CFG 3. That makes the speed-up a practical alternative rather than just a lower-quality preview mode.
Key Capabilities
- The 1024-pixel fine-tune uses the MOSS-Image architecture and its existing inference stack, so it can substitute for MOSS-Image-SFT in compatible workflows.
- Two distilled LoRAs target 8-step sampling; the DMD2 version adds a GAN-based training stage and achieves the strongest reported PartiPrompts result.
- BF16 and INT8 ConvRot checkpoints offer different memory and precision trade-offs, with native ComfyUI support alongside Diffusers.
- The published checkpoint is an EMA snapshot from a fine-tune trained on 704,000 images, rather than a new architecture or text encoder.
Who It’s For
Great fit if you already use MOSS-Image workflows and want a higher-scoring fine-tune or a fast 8-step generation path. Look elsewhere if you need a general-purpose model with broad language coverage: the model is configured for English prompts, and the reported benchmark is a 100-prompt CLIP-based comparison, not a comprehensive measure of image quality.
