Lodestones' Kromoss Slashes MOSS-Image Generation From 50 to 8 Steps
Lodestones ships Kromoss, a full fine-tune of OpenMOSS's 5B diffusion transformer at 1024px, bundled with 8-step turbo LoRAs that beat the base model on PartiPrompts.
- Kromoss is a full fine-tune of MOSS-Image-Base, a 5B text-to-image diffusion transformer, at 1024px.
- Ships two 8-step turbo LoRAs distilled via DMD and DMD2, running at CFG 1 with no negative branch.
- Scores 28.13 CLIP on PartiPrompts with the DMD2 LoRA, beating base MOSS-Image's 27.19.
- Drop-in replacement for MOSS-Image-SFT in both Diffusers and ComfyUI, with int8 ConvRot build at 5.4 GB.
- Trained on 704k images with hybrid Muon-AdamW, lr 9e-6, EMA 0.999, following the MOSS-Image final-stage recipe.
- Apache 2.0 licensed, available at huggingface.co/lodestones/Kromoss.
Kromoss, a community checkpoint from lodestones, has climbed the Hugging Face trending chart after adding two distilled turbo adapters to OpenMOSS’s 5-billion-parameter image model. The release is a full-parameter fine-tune of MOSS-Image-Base, and its LoRAs reduce the default sampling path from 50 steps to eight.
Developers already using the MOSS-Image code can replace the SFT checkpoint without changing the pipeline architecture. Kromoss retains the same text encoder, VAE, prompt format, JSON layout controls, and ComfyUI graph structure.
A 5B base built for text and layouts
OpenMOSS released MOSS-Image-Base and MOSS-Image-SFT on October 16, 2026. The Apache 2.0 release uses a diffusion transformer, which generates images by repeatedly removing noise under the guidance of a text prompt. It supports prompts of up to 4,000 tokens, English and Chinese text rendering, photorealistic and anime styles, and JSON prompts that specify objects, text, and spatial positions.
Kromoss preserves that architecture and uses the Qwen3.5-9B text encoder with the FLUX.2 small VAE decoder. The 5B label covers the image transformer alone; the complete runtime also loads the 9B text encoder, VAE, and intermediate activations, so total memory requirements exceed the transformer’s parameter footprint.
| Component | Kromoss configuration | Developer impact |
|---|---|---|
| Image model | MOSS-Image 5B transformer | Compatible with the existing MOSS pipeline |
| Text encoder | Qwen3.5-9B | Supports long prompts but adds substantial memory use |
| Decoder | FLUX.2 small VAE | Runs in float32 in the recommended configuration |
| Layout input | Natural language and JSON | Existing structured-layout prompts remain usable |
| Default sampling | 50 steps, CFG 3 | Uses a layer-skipped unconditional pass |
| Turbo sampling | Eight steps, CFG 1 | Requires one of the unmerged LoRA adapters |
What the reported score covers
The release reports results from a 100-prompt PartiPrompts evaluation using CLIP ViT-L/14. CLIP scores estimate how closely generated images align with prompt semantics, with higher values indicating stronger measured alignment.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
