<h1 id="moss-image-or-forge-neo-extension"><strong>MOSS-Image | Forge Neo Extension</strong></h1><p>Generate with <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">MOSS-Image</a> inside <a target="_blank" rel="ugc" href="https://github.com/Haoming02/sd-webui-forge-classic/tree/neo">Forge Neo</a>, using the checkpoint, preset and Generate button you already know: text-to-image with sharp text in English and Chinese, image-to-image, and layouts placed with boxes in the prompt.</p><p>Requires an up-to-date <strong>Forge Neo</strong> (the <code>neo</code> branch). On an older version it stays disabled and says so in the console, and Forge Neo keeps working as usual.</p><hr /><h1 id="features"><strong>Features 🚀</strong></h1><h3 id="native-generation">🖼️ Native Generation</h3><ul><li><p>Runs inside txt2img and img2img: no separate tab, no extra environment</p></li><li><p>The SFT and Base models with their Qwen3.5 9B text encoder, through the <strong>VAE / Text Encoder</strong> selector you already use</p></li><li><p>Uses the same autoencoder as FLUX.2 and FLUX 3, so if you have <code>flux2-vae</code> you already have it</p></li><li><p>About a minute per 1024x1024 image at 50 steps on an A40 with the INT8 files</p></li></ul><h3 id="text-and-posters">📝 Text and Posters</h3><ul><li><p>Titles, captions and signs come out as written, in English and in Chinese: put the exact words in quotes</p></li><li><p>Any size that is a multiple of 16, such as 832x1248 for a poster</p></li></ul><h3 id="layout-by-boxes">📐 Layout by Boxes</h3><ul><li><p>Write the prompt as a JSON layout and tell the model where each thing goes, with a <code>bbox</code> per item, in the format of the MOSS-Image user guide</p></li><li><p>Works with the Base and the SFT model</p></li></ul><h3 id="image-to-image">🔁 Image-to-Image</h3><ul><li><p>Put a picture in img2img and describe the result: around 0.6 denoising keeps the scene and follows the details you change</p></li></ul><h3 id="8-step-lora">⚡ 8-Step LoRA</h3><ul><li><p>lodestones' 8-step LoRA works with the SFT model: pictures in 8 steps instead of 50, with no negative prompt pass</p></li></ul><h3 id="upstream's-guidance">🧭 Upstream's Guidance</h3><ul><li><p>The negative pass skips block 22 as the official pipeline does, with a setting to change it</p></li></ul><h3 id="safe-by-design">🛡️ Safe by Design</h3><ul><li><p>Changes no Forge Neo file: remove the extension and everything is as it was</p></li><li><p>Checks your setup at startup and tells you which file is missing instead of failing with a generic error</p></li><li><p>Other models are not affected</p></li></ul><hr /><h1 id="recommended-settings"><strong>Recommended Settings ⚙️</strong></h1><ul><li><p><strong>Euler</strong>, schedule type <strong>MOSS</strong>, 50 steps, CFG 3. The <strong>moss</strong> UI preset sets these values</p></li><li><p>Leave the negative prompt empty with the SFT model. Use sides that are a multiple of 16</p></li></ul><h1 id="installation"><strong>Installation 📦</strong></h1><ol><li><p>Open Forge Neo WebUI and go to <strong>Extensions</strong> → <strong>Install from URL</strong></p></li><li><p>Paste: <code>https://github.com/eduardoabreu81/moss-image-forge-neo</code></p></li><li><p>Click <strong>Install</strong> and restart the WebUI</p></li><li><p>From <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image">Comfy-Org/MOSS-Image</a>, place the INT8 checkpoint (SFT or Base), the Qwen3.5 9B text encoder and <code>flux2-vae</code> in the usual folders, then restart</p></li><li><p>Pick the <strong>moss</strong> UI preset, the checkpoint, and the text encoder and autoencoder under <strong>VAE / Text Encoder</strong></p></li></ol><hr /><p><strong>Documentation:</strong> <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo#-installation">File list</a> · <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo#-whats-new">What's new</a> · <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo">GitHub</a> · License: AGPL-3.0. The MOSS-Image weights are released by the OpenMOSS Team under Apache-2.0 and are not included.</p>

MOSS-Image | Forge Neo Extension

<p>v0.1.1: First release on Civitai</p><ul><li><p><strong>MOSS-Image in Forge Neo:</strong> the SFT and Base models with their Qwen3.5 9B text encoder, with sharp text in English and Chinese.</p></li><li><p><strong>Image-to-image and layouts by boxes:</strong> put things where you want them with a JSON layout in the prompt.</p></li><li><p><strong>8-step LoRA:</strong> lodestones' LoRA works with the SFT model, pictures in 8 steps instead of 50.</p></li><li><p><strong>Plays well with other extensions:</strong> the extension is no longer disabled when another one, such as FLUX 3, hooks the loader first. <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo/issues/1">Issue #1</a></p></li></ul><p>Full changelog: <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo#-whats-new">README</a> · Source: <a target="_blank" rel="ugc" href="https://github.com/eduardoabreu81/moss-image-forge-neo">GitHub</a></p>

Files with SHA256: 052a6066d5e8f225d1bf8e00df5b5d489ac8cd9e40c89e25fa8ab06921f0d010