What Is Kromoss? MOSS-Image Community Fine-Tune: vs SFT & ComfyUI Guide

Quick Answer

Kromoss is a third-party full fine-tune of MOSS-Image-Base; sharper than Base, drop-in SFT replacement.

Less than a week after MOSS-Image went open-source, the community shipped a fine-tune: on October 19, 2026, lodestones released Kromoss — a full fine-tune of MOSS-Image-Base aimed at fixing the Base model's blurry outputs. Note: this is third-party community work, not an official OpenMOSS release. This guide covers who made it, how it differs from official SFT and the 8-step distilled LoRA, and how to run it in ComfyUI.

1. What Kromoss Is (30-Second Brief)

Kromoss is a model released by lodestones on October 19, 2026 (source: HuggingFace page https://huggingface.co/lodestones/Kromoss, verified 2026-10-21; also mirrored on Civitai at 5/5 from 101 reviews). It is a full fine-tune of MOSS-Image-Base at 1024px, trained on 704k images, currently at checkpoint update 11000 (EMA weights) — and still training, with the page updated regularly. Its pitch in one line: fixing the Base model's blurriness. The architecture, inference pipeline, text encoder (Qwen3.5-9B) and VAE (FLUX.2 small decoder) are identical to MOSS-Image-SFT, and the repo layout matches, so it is a drop-in replacement for SFT — whatever workflow you run with SFT works unchanged with Kromoss. Two things to be clear about upfront: first, Kromoss is third-party community work, not an official OpenMOSS release (the page says 'Built on MOSS-Image by the OpenMOSS Team', i.e. built on top of the official model). Second, lodestones is also behind the Chroma series and the MOSS-Image 8-step distillation, with a solid reputation in fast model iteration — but that does not change its third-party status.

2. Kromoss vs SFT vs Distilled: How to Choose

The relationships: SFT is the official supervised fine-tune of Base (the officially recommended generation build); Kromoss is a third-party full fine-tune of Base (aiming for sharper outputs than Base); the 8-step distilled LoRA is a third-party distillation of SFT's 50-step sampling down to 8 steps (aiming for speed). They solve different problems. Choose by need: most stable official renders — SFT; sharper Base-family outputs — try Kromoss (drops straight into SFT's slot); speed and batch drafting — the 8-step distilled LoRA. Note the distilled LoRA stacks on SFT, while Kromoss itself can also take the distilled LoRAs (the author ships two 8-step turbo LoRAs, with the dmd2 build scoring best on PartiPrompts). One common confusion: lodestones' 8-step distilled LoRA only supports SFT, not Base. But since Kromoss shares SFT's architecture, the author confirms both turbo LoRAs transfer to Kromoss. Scores are in section 4 — with the caveat that they are author-measured, marked preliminary.

3. ComfyUI Setup: Download & Load

Current ComfyUI master has native MOSS-Image support — no custom nodes needed (source: HF page ComfyUI section, verified 2026-10-21). Kromoss single files follow the same layout as Comfy-Org's MOSS-Image files: 1) Diffusion model (pick one): kromoss_bf16.safetensors goes in models/diffusion_models/, loaded with Load Diffusion Model. For lower VRAM, use kromoss_int8_convrot.safetensors (int8, 5.4 GB, ConvRot W8A8 quantized). 2) Text encoder: qwen3.5_9b_bf16.safetensors goes in models/text_encoders/, with the CLIP Loader type set to moss_image. 3) VAE: flux2-vae_small.safetensors goes in models/vae/, loaded with Load VAE. 4) Sampling: full model at 50 steps, CFG 3, flow-matching Euler. To speed up, put 8-steps-dmd2-lora.safetensors (recommended) or 8-steps-dmd-lora.safetensors in models/loras/, load with LoraLoaderModelOnly at strength 1, then set KSampler to euler/simple, 8 steps, CFG 1. Do not merge the LoRA into the bf16 weights (the author explicitly warns the delta is close to bf16's rounding step). diffusers users can use the apply_lora snippet from the page to load it dynamically. kromoss_bf16.safetensors drops in for moss_image_sft_bf16.safetensors in Comfy-Org workflows.

4. Reading the Scores (Author-Measured, Preliminary)

The author published PartiPrompts scores on the HF page (100 prompts, CLIP ViT-L/14, measured at update 11000): MOSS-Image-Base at 50 steps 27.19; Base + dmd-lora at 8 steps 27.53; Base + dmd2-lora at 8 steps 27.59; Kromoss at 50 steps 27.90; Kromoss + dmd-lora at 8 steps 27.99; Kromoss + dmd2-lora at 8 steps 28.13 (best). How to read this: first, these are author-measured, not independent benchmarks — marked preliminary. Second, they were measured on a mid-training checkpoint (update 11000) and the model is still training, so numbers will move. Third, the trend favors Kromoss across the board, with the dmd2 LoRA on top — consistent with the 'fix the blur' positioning. But CLIP score only captures text-image alignment, not aesthetics or fine detail; check the side-by-side sample grids on the page for yourself.

5. Commercial Use: Apache-2.0

The HF page states the weights and bundled inference code (pipeline_moss_image.py, derived from MOSS-Image-SFT) are released under Apache License 2.0 — commercial use allowed, no revenue cap. This matches MOSS-Image's own open-source terms. Note this covers the model weights; ownership of images you generate with it follows the usual understanding (yours), but skim the full Apache-2.0 text before commercial use (mainly the attribution and patent-grant parts).

6. Traction & Update Pace

As of 2026-10-21, the HuggingFace page shows 11,608 downloads and 196 likes (HF API data); the Civitai mirror sits at 5/5 from 101 reviews, with active community discussion (multiple hands-on threads on vibingtalk and similar forums). The page was last updated on 2026-10-20 — the author is still training, and checkpoints move forward regularly. Check the HF page's latest notes before use: what is update 11000 today may not be the latest tomorrow.

7. Sources & Disclaimer

Sources: 1) https://huggingface.co/lodestones/Kromoss (page verified 2026-10-21): model description, ComfyUI usage, PartiPrompts scores, training details, Apache-2.0 license; 2) HuggingFace API (2026-10-21): 11,608 downloads, 196 likes, last updated 2026-10-20; 3) https://civitai.com/models/3009368/kromoss (third-party mirror): 5/5 from 101 reviews, published 2026-10-19; 4) vibingtalk.com (Oct 2026, third-party community discussion): hands-on feedback and update threads. Disclaimer: Kromoss is third-party community work, unrelated to OpenMOSS; PartiPrompts scores are author-measured (preliminary), not independent benchmarks; this site is not affiliated with lodestones or OpenMOSS; this article is informational only, and all model rights belong to their owners.

Kromoss vs Official SFT vs 8-Step Distilled LoRA

Dimension	Kromoss (3rd-party fine-tune)	SFT Original (official)	8-Step Distilled LoRA (3rd-party)
Maker	lodestones (community), still training	OpenMOSS official	lodestones (community), WIP
Form	Full weights (Base full fine-tune)	Full weights (supervised fine-tune)	LoRA, stacked on SFT
Solves	Fixes Base blurriness	Official recommended generation build	Speeds 50 steps to 8
Sampling	50 steps CFG 3; 8 steps with LoRA	50 steps, CFG 3	8 steps, no CFG
Interchangeable	Same arch as SFT, drop-in	Baseline	SFT only, not Base
License	Apache 2.0, commercial OK	Apache 2.0, commercial OK	Apache 2.0 (follows base)
Use for	Sharper Base-family outputs	Final renders, commercial delivery	Fast drafts, batch ideation

Frequently Asked Questions

Is Kromoss an official release?

No. It is a community fine-tune by third-party lodestones on HuggingFace, built on OpenMOSS's MOSS-Image-Base. OpenMOSS currently ships only Base and SFT. Check the HF page's latest notes before use — the model is still training and updates frequently.

How does Kromoss relate to the 8-step distilled version?

They are different things: Kromoss is full weights (a full fine-tune of Base) that replaces SFT directly; the 8-step distilled build is a LoRA stacked on SFT (or Kromoss) to cut 50 steps to 8. The author's two turbo LoRAs (dmd/dmd2) both transfer to Kromoss, with dmd2 scoring highest.

What GPU do I need to run Kromoss locally?

Same architecture as SFT, so similar VRAM needs: the bf16 full build wants a 24GB-class card; for less VRAM use the int8 quantized build (5.4GB file) with text-encoder offloading — workable on ~12GB cards. 1024x1024 is the baseline resolution.

Can I trust Kromoss's scores?

The PartiPrompts numbers on the page are author-measured (update 11000 checkpoint, model still training) — marked preliminary, for reference only. CLIP score captures text-image alignment only. Look at the sample comparison grids on the page and community hands-on threads alongside the numbers.