README.md

metadata

title: MOSS-Image Workflow
emoji: 🖼
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.30.0
python_version: '3.12'
app_file: app.py
hf_oauth: true
hf_oauth_scopes:
  - inference-api
  - write-repos
pinned: false
license: apache-2.0
tags:
  - ml-intern
short_description: MOSS-Image text-to-image as a gr.Workflow node canvas.

MOSS-Image Workflow

                   

A Gradio Workflow (node-canvas) version of the official OpenMOSS-Team/MOSS-Image demo, built on gr.Workflow in gradio 6.30.0, with the App view styled after the MOSS-Image landing page.

Feature parity with the official demo

Every sampling control of the original Space is a reference node wired into the MOSS-Image Generate function node:

Original demo	Here
Model dropdown (Base / SFT)	Model reference (base or sft); both checkpoints loaded at startup
Prompt / Negative Prompt	Prompt / Negative Prompt references (same default text)
Width × Height (presets)	Width / Height references (multiples of 16)
Steps / CFG / Shift / Seed	Steps / CFG / Shift / Seed references (-1 = random)
Images per prompt	Images reference → gallery output
Max text length	Max Text Length reference (1–4096)
Skip-layer modes (Off / Negative-branch skipping / Standard SLG)	Skip Layers (22 = negative-branch skipping, empty = off), SLG Layers, SLG Scale, Skip Start, Skip Stop references
Example gallery with one-click restore	Examples strip under the hero: the original Space's 20 showcase images, hot-linked; clicking one fetches its parameters.json and restores prompt, model, size, sampling and SLG settings into the form
—	Turbo mode: set Model to turbo for the 4-step distilled LoRA (see below)

Turbo (4-step) mode

Set the Model field to turbo to run the 4-step distilled LoRA from akhaliq/MOSS-Image-SFT-4-steps (distilled via data-space/Euler-native DMD, arXiv:2610.12102) — roughly 10x faster per image than the 50-step sampler.

• The fp32 LoRA is wrapped on the SFT transformer at startup, unmerged (turbo_lora.py, vendored from the model repo), and toggled per request by its alpha — so Base and SFT generations are unchanged, and no third transformer is loaded into GPU memory.
• Turbo always runs at CFG 1; the negative prompt is ignored.
• Left at the default 50, Steps auto-maps by resolution per the model card: 4 up to 1024px, 6 at ~1536px, 8 at 2048px. Set Steps yourself to override.

Not ported: showcase add/delete with live refresh (the Workflow App view is generated from the graph and has no upload-to-gallery mechanism), and ZeroGPU AOT compilation (this app runs the eager backend — set it up like the original if you need the extra speed).

Design

The App view is auto-generated by gr.Workflow, so the landing-page look is applied through launch(theme=…, css=…, head=…): Inter type, a light gradient canvas, the official MOSS-Image lockup (light/dark aware) above the "Text-to-Image Generation with Native Layout Control" headline, a top nav with model/code links, and a Soft blue theme. The canvas view is untouched.

Running it

Generation runs on GPU via @spaces.GPU(duration=80) per node call, so this Space needs ZeroGPU hardware (Settings → Hardware). hf_oauth: true is enabled so the owner can edit the canvas and signed-in visitors spend their own ZeroGPU quota; anonymous visitors fall back to the small shared allowance.

Local development:

pip install -r requirements.txt
python app.py                           # full run, needs a CUDA GPU
MOSS_WORKFLOW_PREVIEW=1 python app.py   # canvas + styling only, no model download

The bundled pipeline_moss_image.py is copied verbatim from the original Space (BSD 3-Clause); model weights are Apache-2.0. The lockup artwork is hot-linked from the original Space's assets. The form is prefilled with showcase example 007 (the photorealistic Kazusa at a Tokyo crosswalk: 1184×1776, 50 steps, CFG 3, shift 5, skip layer 22, seed 3537712559) so visitors can just hit Run and reproduce it exactly. Other defaults match the official demo: Euler sampling and negative-branch skipping.

API example

from gradio_client import Client

client = Client("akhaliq/moss-image-workflow")
client.view_api()  # exact parameters of /generated_images