Kromoss: MOSS-Image fine-tuned, with Turbo LoRAs for 8-step images

What this is about

Hugging Face user lodestones has released a text-to-image model called Kromoss. It is a full fine-tune (all transformer parameters are trained further) of MOSS-Image-Base from the OpenMOSS team, at 1024 px resolution. The architecture, inference pipeline, text encoder (Qwen3.5-9B) and VAE (the "small decoder" from FLUX.2, which turns the compressed latent format back into images) are identical to MOSS-Image-SFT. That means you can use Kromoss as a direct replacement. The model has about 5.1 billion parameters and is released under Apache-2.0.

What's in the package

Besides the main model, the repo contains two "Turbo" LoRAs (LoRAs are small add-on weights that change one aspect of a model's behavior). They were distilled from MOSS-Image-Base, meaning they were trained to imitate the output of a slow process (50 steps, CFG 3) in just 8 steps:

• 8-steps-dmd-lora.safetensors: data-space distillation with DMD, 12,000 updates.
• 8-steps-dmd2-lora.safetensors (recommended): the same run, continued for 3,000 more updates with the second stage of DMD2, a GAN discriminator against real images, at a quarter of the learning rate. According to the card, this gives slightly more contrast and detail.

With a LoRA you need CFG 1 (CFG, classifier-free guidance, controls how closely the image follows the prompt). There is then no negative branch, which halves the compute per step. The card says 4 steps also work. Both LoRAs are fp32 with rank 128 and alpha 128 on all projections in every block. Use only one at a time, and don't merge them into the bf16 weights, because the change sits close to bf16's rounding step.

Training

Training was a fine-tune of all transformer parameters at 1024 px; images below 896² pixels were discarded. It used the final-stage recipe from the MOSS-Image report, scaled to 64 images per update: hybrid Muon-AdamW (the optimizer) with a learning rate of 9e-6 and 200 warmup steps, truncated logit-normal timesteps, q/k-norm decay in the last block down to one, and EMA 0.999. The released weights are the EMA weights from update 11000, trained on 704,000 images.

The numbers

The CLIP ViT-L/14 score was measured on 100 PartiPrompts (higher is better; CLIP measures how well image and text match):

The differences are small and rest on only 100 prompts and an automated metric. Whether the images look better to people is backed only by side-by-side example images on the card, not by a user study.

What you need to try it

• GPU: The card gives no VRAM requirement. The transformer has 5.1B parameters (about 10 GB in bf16), plus the Qwen3.5-9B text encoder (about 18 GB in bf16). This is our own estimate; CPU offloading or the Int8 version makes it more manageable.
• Int8 variant: kromoss_int8_convrot.safetensors is 5.4 GB (W8A8 with Hadamard rotation) and works in ComfyUI and with the LoRAs.
• Software: diffusers ≥ 0.38, transformers ≥ 5.3, accelerate ≥ 1.10, or a current ComfyUI (master).
• Mac/CPU/Colab: Not mentioned in the source. The examples use CUDA.

Using it in ComfyUI and diffusers

In ComfyUI, the files go in the usual folders (diffusion_models, text_encoders, vae, loras). Load the text encoder with type moss_image. With a LoRA: KSampler, euler/simple, 8 steps, CFG 1. Without a LoRA: 50 steps, CFG 3, shift 5, and layer 22 is skipped in the negative branch. Included scripts convert your own MOSS-Image fine-tunes to ComfyUI format (convert_to_comfy.py) and quantize them to Int8 (quantize_convrot.py); the author says they reproduce the Comfy-Org files byte for byte. In diffusers, you load the pipeline with trust_remote_code=True, and the LoRA is applied unmerged via a forward hook.

Our take

Kromoss is a community fine-tune, not a major release. It's interesting if you already use MOSS-Image and want a slightly better model or faster generation with little effort. The quality gains rest on a small CLIP test alone, but the Turbo LoRAs are a practical win: 8 steps with no negative branch take far less compute than 50 steps with CFG. Still unclear are the training data (only the figure of 704k images is given), how it compares with other image models, and how it handles text rendering or styles. The model card is in English.