# MOSS-Image SFT - New 4-Step Distillation LoRA, using eDMD + LADD - producing high detail images (T2I, T2I with Layout Boxes, I2I Edits) — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/dc77289c552d9d4c

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-19 11:00 ingested 2026-11-20 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1y06n27/mossimage_sft_new_4step_distillation_lora_using/)

## retained text



```
This is [my personal take on MOSS Distillation, creating a new 4 Step LoRA](https://huggingface.co/lvladikov/MOSS-SFT-Distill-4step-LoRA) using a mix of eDMD and LADD (latent adversarial diffusion distillation) as training method.

I am fully aware that there are existing Step Distillation LoRAs for MOSS, and I have been using them for some time (like those from [akhaliq](https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps), which are very good), but I wanted to step up the game a bit, and make LoRA that produces higher detail images, and hopefully I've achieved that. I have trained on a lot more resolutions, and image ratios, and also used not only teacher input but also real photo crops from public high resolution datasets (I did the same with [my Krea 2 Turbo LoRAs](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA), for those that have been following). I have provided **workflows for ComfyUI Local and Cloud - for Text 2 Image, Text 2 Image with Layouts/Boxes and Image 2 Image** (edits). The LoRA has been trained and optimised to **work well with Text** too.

Quick summary:

* ⚡ **4 steps, CFG 1** — 4 model calls against the teacher's 100 (50 steps, two calls each at CFG 3), on MOSS's own 4-step schedule
* ⏱️ **About 14× less GPU time** — 26.6 s against 387.3 s for the same 22 pictures at 1024×1024 on Comfy Cloud
* 🎯 **Close to the teacher's tone, with more fine detail** — on 21 test prompts at 1824×1024 the mean lightness sits 0.7 L\* above the teacher's (scale 0–100), haze 0.94×, colour 1.02×; fine texture **1.22×** the teacher's on 22
* 🔤 **Writes the words a prompt spells out** — every requested string right on 14 of the 15 text test prompts at 1024×1024 (the teacher 15, [akhaliq's 4-step LoRA](https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps) 12)
* 📐 **12 trained resolutions** — wide, tall and square, from 768×512 and 512×768 up to 1824×1024 and 1024×1824
* 🧩 **Eight ComfyUI workflows** — text to image, layout, image edit and optional PiD 4× (with supersampled latent for better results), each for local ComfyUI and for Comfy Cloud
* 🎲 **34,000 prompts** from [Lakonik's 3M-prompt dataset](https://huggingface.co/datasets/Lakonik/t2i-prompts-3m) and **2,877 prompts that spell out words** — signs, menus, posters, labels
* 📷 **About 8,000 captioned photo crops and 3,600 teacher pictures**
* 🔢 **15,000 training samples** in the shipped weights — rank 64, 252 modules
* 📅 Trained on a single **NVIDIA B300** in about 30 hours

Details, Workflows and of course the LoRA (in ComfyUI and Diffusers formats) at my repo: [https://huggingface.co/lvladikov/MOSS-SFT-Distill-4step-LoRA](https://huggingface.co/lvladikov/MOSS-SFT-Distill-4step-LoRA)

While B300 is an amazing and efficient card to train on (in the cloud), it is also very expensive, so I will continue for a short while more (current released checkpoint is 15K), and an update may be published in few days, but won't train for too long (the cost reasons). So watch out for possibly one more update in the days to come.

As I've mentioned on the model card, I highly recommend using also the community **MOSS Realism Enhancer** LoRA which can be added beside mine — [Civitai](https://civitai.red/models/3023438?modelVersionId=3433342) · [Hugging Face copy](https://huggingface.co/TalasCloud/MOSS-Realism-Enhancer-v1-BF16-LoRA). It is a separate LoRA by its own author under its own licence, not part of this training, and none of the pictures displayed in this post use it. In ComfyUI it goes in a second LoRA loader after this one. Its is very good for even more photorealistic skin.

Enjoy!
```



document fetched 2026-11-20 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1y06n27/mossimage_sft_new_4step_distillation_lora_using
