Writes long English and Chinese text exactly as you quote it and places titles and objects where your JSON layout says. Uncensored, in an 8-step or a 50-step preset.
No GPU. No setup. Cancel anytime.
Quote the words and MOSS-Image writes them: whole whiteboards, bilingual menus and concert-poster small print. Click any image to open it at full size.
Three sections of handwritten notes, every word from the prompt.
Four drinks with English names, Chinese names and prices on one chalkboard.
Headline, date, venue, three band names and the small print.
A beautiful young woman wearing a white T-shirt with a blue "MOSS" logo smiles at the camera, holding a black marker. Behind her, a glass whiteboard is covered in neat handwritten black marker text that reads: "1. MOSS-Image under the hood: a 5B single-stream DiT model, trained with Muon. 2. What MOSS-Image is great at: a) Complex text rendering in English and Chinese, with automatic layout. b) JSON layouts put titles and objects exactly where you want them. 3. MOSS-Image on Sogni: the full 50-step SFT or the fast 8-step preset. Personal LoRAs supported."Use prompt →
MOSS-Image SFT is OpenMOSS’s 5-billion-parameter text-to-image model. Describe a scene in English or Chinese, use tag prompts, or supply an experimental JSON layout to guide object placement and writing.
Choose the 8-step preset for fast drafts, or standard SFT at 50 steps for the finest detail. Both start at 1024 × 1024.
Choose catalog LoRAs or import compatible Personal LoRAs through My LoRAs, and use them on either preset. The 8-step preset applies personal adapters after its built-in DMD adapter.
Both MOSS presets are uncensored, supporting private mature-theme creativity alongside everyday image generation.
Both presets generate from text and accept compatible Personal LoRAs. Select 8-Step DMD for quick iteration or Standard SFT for the full sampling recipe.
| Variant | Details | From | |
|---|---|---|---|
Standard SFT · 50 steps Qualitymoss_image_sft_int8_convrot | Input: Captions · tags · experimental JSON layouts Resolution: 512–2048 px · multiples of 16 Steps: 50 default · 50–100 Guidance: 3 default · 3–6 | 8.86 Spark · $0.044 | Create → |
8-Step DMD Fastmoss_image_sft_dmd8_int8_convrot | Input: Captions · tags · experimental JSON layouts Resolution: 512–2048 px · multiples of 16 Steps: 8 · fixed | 1.25 Spark · $0.0063 | Create → |
Get the most out of MOSS-Image by covering:
| You provide | You get |
|---|---|
| Words in quotes | Those words written as you quote them, in English or Chinese, from one-word labels to whole whiteboards. |
| A JSON layout | Each title, object or comic panel placed in the box you gave it. See Place every element. |
| Any aspect ratio | 512 to 2048 px per side in multiples of 16, from 2048 × 512 banners to tall posters. |
| A Personal LoRA | Your own style or character on either preset, from My LoRAs. |
MOSS-Image makes images from text. It does not edit an uploaded photo or follow reference images; for that, choose an image-editing model such as FLUX 3 Image Klein.
For exact composition, write the prompt as a JSON layout: a summary describing the whole image and an items list where each title, object or panel has a caption and a box. Boxes are [top, left, bottom, right] in pixels measured from the image’s center. Below, the same five elements in two arrangements, with the same seed. The colored outlines are the boxes from each prompt.
{
"summary": {
"meta": [
"poster",
"product campaign"
],
"canvas": "single",
"caption": "A bright summer soda campaign poster on a sunny yellow background with soft natural shadows. Clean, bold modern sans-serif typography in deep navy. Fresh, crisp, playful."
},
"items": [
{
"label": "text",
"caption": "Huge bold deep-navy sans-serif headline on one line.",
"text": "SUNNY SODA",
"bbox": [
-448,
-448,
-320,
448
]
},
{
"label": "text",
"caption": "Smaller deep-navy sans-serif tagline on one line.",
"text": "Squeezed today. Gone by noon.",
"bbox": [
-304,
-352,
-240,
352
]
},
{
"label": "object",
"caption": "A tall clear glass bottle of sparkling lemon soda with cold condensation, a yellow cap and a white paper label, standing upright.",
"bbox": [
-192,
-416,
448,
-64
]
},
{
"label": "object",
"caption": "A small pile of fresh lemons, one cut in half showing its juicy inside, with two green leaves.",
"bbox": [
192,
32,
448,
448
]
},
{
"label": "text",
"caption": "A round white sticker badge with deep-navy lettering on two lines.",
"text": "ONLY\n$3",
"bbox": [
-160,
160,
64,
384
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
} Use prompt →
{
"summary": {
"meta": [
"poster",
"product campaign"
],
"canvas": "single",
"caption": "A bright summer soda campaign poster on a sunny yellow background with soft natural shadows. Clean, bold modern sans-serif typography in deep navy. Fresh, crisp, playful."
},
"items": [
{
"label": "text",
"caption": "Huge bold deep-navy sans-serif headline on one line.",
"text": "SUNNY SODA",
"bbox": [
320,
-448,
448,
448
]
},
{
"label": "text",
"caption": "Smaller deep-navy sans-serif tagline on one line.",
"text": "Squeezed today. Gone by noon.",
"bbox": [
240,
-352,
304,
352
]
},
{
"label": "object",
"caption": "A tall clear glass bottle of sparkling lemon soda with cold condensation, a yellow cap and a white paper label, standing upright.",
"bbox": [
-448,
64,
192,
416
]
},
{
"label": "object",
"caption": "A small pile of fresh lemons, one cut in half showing its juicy inside, with two green leaves.",
"bbox": [
-64,
-448,
192,
-32
]
},
{
"label": "text",
"caption": "A round white sticker badge with deep-navy lettering on two lines.",
"text": "ONLY\n$3",
"bbox": [
-448,
-416,
-224,
-192
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
} Use prompt → Layouts also hold multi-panel stories and dense information graphics.
{
"summary": {
"meta": [
"comic"
],
"canvas": "multi_panel",
"caption": "A cheerful three-panel comic strip read left to right, equal white gutters and black panel borders, flat bright cartoon style. The same small round white robot with glowing blue eyes appears in every panel: it finds a seed, waters a sprout in a terracotta pot, then celebrates beside a tall sunflower."
},
"items": [
{
"label": "panel",
"caption": "Left panel: the small round white robot kneels in a sunny garden, holding up one seed and looking at it curiously.",
"bbox": [
-368,
-656,
368,
-240
]
},
{
"label": "text",
"caption": "A white comic speech bubble with a black outline and bold black comic lettering.",
"text": "WHAT IS THIS?",
"bbox": [
-336,
-608,
-192,
-288
]
},
{
"label": "panel",
"caption": "Middle panel: the same small round white robot waters a tiny green sprout in a terracotta pot with a blue watering can.",
"bbox": [
-368,
-208,
368,
208
]
},
{
"label": "text",
"caption": "A white comic speech bubble with a black outline and bold black comic lettering.",
"text": "GROW, LITTLE ONE.",
"bbox": [
-336,
-176,
-192,
176
]
},
{
"label": "panel",
"caption": "Right panel: the same small round white robot throws its arms up in joy beside a tall sunflower growing from the terracotta pot.",
"bbox": [
-368,
240,
368,
656
]
},
{
"label": "text",
"caption": "A white comic speech bubble with a black outline and bold black comic lettering.",
"text": "WE DID IT!",
"bbox": [
-336,
288,
-192,
608
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
} Use prompt →
{
"summary": {
"meta": [
"infographic",
"educational poster"
],
"canvas": "single",
"caption": "A clean, friendly educational infographic about honeybees on a warm cream background, honey-yellow and charcoal palette, flat vector illustration, rounded sans-serif typography, thin dotted connector lines from the fact cards to the bee."
},
"items": [
{
"label": "text",
"caption": "Huge bold charcoal headline centered at the top, on one line.",
"text": "THE LIFE OF A HONEYBEE",
"bbox": [
-736,
-480,
-616,
480
]
},
{
"label": "text",
"caption": "A smaller honey-yellow subtitle centered under the headline.",
"text": "Four facts from the hive",
"bbox": [
-592,
-320,
-536,
320
]
},
{
"label": "object",
"caption": "A large cheerful flat vector illustration of a honeybee flying in a three-quarter side view, a round fuzzy yellow body with black stripes, translucent wings and a friendly face, with a few hexagonal honeycomb cells behind it.",
"bbox": [
-496,
-320,
112,
320
]
},
{
"label": "text",
"caption": "A rounded white fact card with a soft shadow, a bold charcoal heading on the first line and one line of smaller charcoal text beneath it.",
"text": "3 BODY PARTS\nHead, thorax and abdomen.",
"bbox": [
160,
-480,
400,
-24
]
},
{
"label": "text",
"caption": "A rounded white fact card with a soft shadow, a bold charcoal heading on the first line and one line of smaller charcoal text beneath it.",
"text": "6 LEGS\nBack legs carry pollen.",
"bbox": [
160,
24,
400,
480
]
},
{
"label": "text",
"caption": "A rounded white fact card with a soft shadow, a bold charcoal heading on the first line and one line of smaller charcoal text beneath it.",
"text": "5 EYES\nTwo compound eyes, three simple.",
"bbox": [
440,
-480,
680,
-24
]
},
{
"label": "text",
"caption": "A rounded white fact card with a soft shadow, a bold charcoal heading on the first line and one line of smaller charcoal text beneath it.",
"text": "4 WINGS\nTwo pairs that hook together.",
"bbox": [
440,
24,
680,
480
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
} Use prompt → OpenMOSS labels layout control experimental. Check each element’s position and every word, and adjust a box or a caption when something lands off.
You don’t have to write the JSON by hand. In the Sogni web app, choose MOSS-Image and Compose by layout: draw a box for each title, object or character, type the words, and Sogni writes this layout for you. See how Layout Composer works.
MOSS-Image 8-Step uses the DMD adapter by lodestones on top of the SFT checkpoint. In our RTX 4090 tests with the weights loaded, it finished 1024 × 1024 images in a median 3.3 seconds; the 50-step recipe took 23.8 seconds. Same prompt and seed below: draft with 8 steps, then render the keeper at 50 for the finest skin, glass and fine print.
In the MOSS-Image technical report, the team measures 96.3 on its own long-text rendering benchmark across English and Chinese, and 0.88 layout accuracy (mIoU) on 7Bench. These are OpenMOSS’s numbers, from its own evaluation.
For more prompting examples, read the OpenMOSS user guide and project page. Developers will find the settings and the layout coordinate format in the MOSS-Image model guide.
Choose a look in Browse LoRAs on either MOSS preset, with sample images to compare before you render.
| Look or control | Catalog LoRAs |
|---|---|
| Film and atmosphere | 2000s Analog Core, Retro Vintage Photo, Fever Dream Mood, Kodachrome Photo |
| Photographic realism | Lenovo UltraReal, GrainScape UltraReal, NiceGirls UltraReal, Canon UltraReal |
| Detail | Moss detail tweaker, MOSS Image MAX ULTRA hires |
| Private mature themes | Mystic X · Sensitive Content Filter off |
Compatible Personal LoRAs from My LoRAs work on both presets too.
Use pay-as-you-go Spark packs for each render (1 Spark = $0.005), or choose a flat-rate Sogni plan for credit-free fair-use generation in the app.
| Variant / configuration | Spark | USD |
|---|---|---|
| Standard SFT · 50 steps · 1024 × 1024 · 50 steps | 8.86 Spark | $0.044 |
| Standard SFT · 50 steps · 2048 × 2048 · 50 steps | 48.4 Spark | $0.24 |
| 8-Step DMD · 1024 × 1024 · 8 steps | 1.25 Spark | $0.0063 |
| 8-Step DMD · 2048 × 2048 · 8 steps | 5.46 Spark | $0.027 |
One Sogni API key reaches every model on the Supernet — call MOSS-Image with the exact model id for the variant you want.
import { SogniClient } from '@sogni-ai/sogni-client';
const client = await SogniClient.createInstance({
appId: crypto.randomUUID(),
apiKey: process.env.SOGNI_API_KEY,
network: 'fast',
});
const project = await client.projects.create({
type: 'image',
modelId: 'moss_image_sft_dmd8_int8_convrot',
positivePrompt: 'a slothicorn surfing a wave of liquid paint, studio lighting',
numberOfMedia: 1,
});
const [url] = await project.waitForCompletion();
console.log(url); // result link — download within 24h curl https://api.sogni.ai/v1/creative-agent/workflows \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $SOGNI_API_KEY" \
-d '{
"input": {
"title": "First render",
"steps": [{
"id": "step1",
"toolName": "generate_image",
"arguments": { "prompt": "a slothicorn surfing a wave of liquid paint, studio lighting", "model": "moss-image-8step" }
}]
},
"confirm_cost": true
}' Example uses moss_image_sft_dmd8_int8_convrot. Use any exact model id from the Variants table above. Full reference at docs.sogni.ai.
Original generations from the standard and 8-step presets. Each card names its preset; Use prompt opens that preset in Sogni Studio. Check small text and fine gradients in your own results.
Use a flat monthly plan for credit-free fair-use generation, or buy Spark packs when pay-as-you-go fits better. Both run on the same creator-owned GPU network.
One flat price in the app. Generate under fair use without a per-image meter.
Image, video, music, and language models in one workspace and one API key.
Prefer pay-as-you-go? Call MOSS-Image by id and pay with Spark packs.
Runs on a decentralized GPU network where workers share subscription revenue.
SFT stands for supervised fine-tuned.
This release accepts text only. Changing a JSON layout produces a new image. For edits that must preserve a photographed person or object, choose a model with reference-image support.
Start at 1024 × 1024 with the 8-Step preset to draft quickly, then render the keeper with Standard SFT at 50 steps. Leave the negative prompt empty initially.
Send the whole JSON layout as the positive prompt. The author’s format uses summary, items, bbox_format yxyx and bbox_coord_space abs_centralized. Boxes are pixel coordinates relative to the canvas center. Layout control is experimental; check the composition and every rendered word.
Choose MOSS-Image 8-Step in Sogni Studio. It uses the lodestones DMD adapter at strength 1 with exactly 8 steps. Its direct model ID is moss_image_sft_dmd8_int8_convrot; the hosted tool selector is moss-image-8step. Check the quote for that preset before rendering.
Import a compatible adapter through My LoRAs, then select the ready entry on either MOSS preset. Personal adapters are shared across the two presets; on 8-Step, they apply after the built-in DMD adapter.
Yes. Both the standard SFT and 8-Step DMD presets are uncensored and support private mature-theme creativity.
Both presets support Lenovo UltraReal, GrainScape UltraReal, NiceGirls UltraReal, Moss detail tweaker, Canon UltraReal, 2000s Analog Core, Retro Vintage Photo, Fever Dream Mood, Kodachrome Photo, MOSS Image MAX ULTRA hires. Mystic X is available with the Sensitive Content Filter off. Open Browse LoRAs in Sogni Create to compare their sample images.
Sogni is built with privacy and creative freedom in mind. Your work remains your own, and inference runs through the Sogni Supernet, a decentralized network of creator GPUs, instead of requiring local hardware or a separate model host. Use is still governed by Sogni's Privacy Policy and Terms of Use.
Create in the app, or build with the API. Your call.