Voxta docs

MOSS-Image

Local image generation with OpenMOSS's MOSS-Image, a 5B diffusion transformer that reads long prompts and renders text accurately. LoRAs, an 8-step Turbo preset and pose control.

MOSS-Image (Apache 2.0) is OpenMOSS's 5B diffusion transformer, run on your own GPU. A Qwen3.5-9B text encoder reads prompts up to 4K tokens, and the FLUX.2 VAE draws the result. It renders words in the image accurately and takes JSON layouts that place text and objects on the canvas: posters, signs, menus, panels.

Experimental. Only an admin can add it. It needs about 10 GB of video memory with the INT8 builds (about 16 GB with bf16), and it is slow without the Turbo preset.

Setup

Add the service

Add Modules → Services → MOSS-Image → Install, then follow the install steps. It installs on its defaults. The Models step offers the checkpoint and text encoder to download now.

Pick a preset

PresetFor
MOSS-Image SFTThe polished model: cleaner text and layouts. 50 steps.
MOSS-Image BaseThe untuned model LoRAs are trained on. Can look more natural and varied.
MOSS-Image SFT Turbo (4 steps)SFT with akhaliq's 4-step Turbo LoRA (~1.2 GB) and CFG 1: the fastest.
MOSS-Image SFT Turbo (8 steps)SFT with lodestones' 8-step Turbo LoRA (~1.2 GB) and CFG 1: an image in seconds.

A Turbo preset's LoRA is fetched into the LoRAs folder the first time the preset runs.

Models

The checkpoint and text encoder are preset settings, so each preset can use its own.

SettingChoices
ModelSFT INT8 (~5 GB, recommended), Base INT8 (~5 GB), SFT bf16 (~9.5 GB), Base bf16 (~9.5 GB). INT8 halves the download and video memory for images close to bf16's. A sharded OpenMOSS release folder also works.
Text EncoderQwen3.5-9B INT8 (~9 GB, recommended) or bf16 (~18 GB). It is streamed through the GPU for each new prompt, keeping as much loaded as the card has room for.
Banding FixSFT can draw faint vertical stripes across skies, fog and walls. Last 5 layers or Last 10 layers (strongest) takes those layers from Base, which also changes the image somewhat. Downloads Base once.

Checkpoints and text encoders live in the module's Models Directory, LoRAs in its LoRAs Directory.

Settings

  • Steps: MOSS-Image is at its best from 50. Fewer is faster and rougher.
  • CFG Scale: 3 suits long prompts, short ones can take up to 6. 1 turns guidance and the negative prompt off, for twice the speed.
  • Timestep Shift: default 5.
  • Negative Branch Skipped Layers: sharpens detail. Default 22. Empty turns it off.
  • Resolution: trained for 1 to 2 megapixels in any shape.
  • Active LoRAs: LoRAs made for MOSS-Image, in the kohya, diffusers or ComfyUI format, from Hugging Face or CivitAI.

Pose control

In the Image playground (Create › Visual › Image), press Pose to pose a 3D figure over the canvas. The image follows it. This uses darask0's control adapter (~0.3 GB), tuned under the preset's Pose settings:

  • Pose Strength: how closely the image follows the pose.
  • Pose Until: the share of the steps the pose acts on. Much more, and the people take on the figure's surface.
  • Pose Looseness: how far the figure's outline is widened. At 0 the people take the figure's exact build.

Prompt assistant

The playground opens the MOSS-Image prompt assistant for this module. Give it an idea and it writes a detailed description, with any words to show in the image quoted exactly. For a poster, a sign or several panels it writes a layout instead. Attach a picture and it writes the prompt that recreates it.

More info

On this page