<h1 id="artius-moss">Artius Moss</h1>A realistic portrait. A floating city. That oddly specific idea you’ve been meaning to try.Artius Moss gives you room to explore them all. It’s a versatile, photorealistic MOSS-Image merge with a carefully balanced fine-tune, LoRAs, and built-in turbo. Create sharp images in 8–10 steps, switch styles, and keep experimenting — even on a GPU with 8 GB VRAM.⚡ 8–10 steps · CFG 1 · ~7 GB model · runs on 8 GB GPUs 📦 Custom text encoder and Flux 2 VAE available on this page<h2 id="one-model.-plenty-of-directions.">🎨 One model. Plenty of directions.</h2>Start with a portrait in soft window light. Try a rainy street glowing with neon. Then build a tiny village on the back of a giant turtle.Artius Moss is made to move between photorealism, illustration, fantasy, fashion, landscapes, and more. The fine-tune and LoRA strengths were tested side by side and adjusted to work together, giving you a broad range of styles without stamping the same look onto every image.Bring a detailed prompt or start with a simple idea. There’s room for both.<h2 id="made-for-experimenting">✨ Made for experimenting</h2><ul><li>Fast from the start. Turbo is already built in. Use 8 steps for quick experiments or try 10 for a little extra detail.</li><li>A new style is a prompt away. Explore realistic photography, painted worlds, character art, or something harder to put a name to — all with the same model.</li><li>Comfortably compact. The optimized INT8 model is around 7 GB and runs in ComfyUI on GPUs with 8 GB VRAM.</li><li>A custom encoder for your prompts. The included Qwen3.5-9B text encoder was prepared specifically for this model. It can also expand a short idea into a more detailed prompt using ComfyUI’s TextGenerate node.</li><li>A workflow you can pick up and use. Drag a showcase image into ComfyUI to load the complete workflow and start exploring.</li></ul><h2 id="start-here">⚙️ Start here</h2>SettingRecommended valueSteps8–10 — start with 8; try 10 for extra detailCFG1 — no negative prompt needed; describe what you want to seeSampler / schedulereuler / simpleResolution1–2 MP, such as 1024×1024 or 912×1632VAEFlux 2 VAE, available in this page’s filesText encoderThe custom encoder from this page, loaded with CLIPLoader → moss_imageThe standard MOSS-Image Qwen3.5-9B encoder also works, but the custom version is recommended.<h2 id="for-the-curious:-what's-inside">🔧 For the curious: what’s inside</h2>A lot of the work happens before you generate your first image.<h3 id="a-merge-that-pays-attention-to-overlap">🧩 A merge that pays attention to overlap</h3>Each source brings something to the model, but some of those contributions overlap. I account for that overlap during merging, with the aim of keeping useful differences while reducing repeated emphasis on what the sources already share.The method is an orthogonalized sum inspired by TSV-Merge, using Löwdin symmetric orthogonalization at full rank.For each tensor, the process:<ol><li>Measures the similarity between the sources’ weight changes, also called deltas.</li><li>Orthogonalizes those deltas to separate their directions before combination.</li><li>Adds the adjusted deltas together using carefully chosen strengths.</li></ol>There’s no low-rank truncation during this merge step. The compact INT8 build comes later, through quantization.<h3 id="full-precision-before-quantization">🔬 Full precision before quantization</h3>All LoRAs are merged in FP32 before quantization, helping preserve fine details during the merge.<h3 id="each-block-gets-individual-attention">⚖️ Each block gets individual attention</h3>The model’s blocks were evaluated for their sensitivity to quantization. Sensitive layers retain full precision; the others use INT8 to keep the model compact while limiting quality loss.<h3 id="the-text-encoder-gets-the-same-treatment">💬 The text encoder gets the same treatment</h3>The custom encoder was calibrated too, keeping its sensitive layers at full precision to preserve prompt understanding in a smaller build.<h2 id="what-you'll-need">📦 What you’ll need</h2><ul><li>An up-to-date ComfyUI installation with MOSS-Image support.</li><li>The Load Diffusion Model node.</li><li>An NVIDIA RTX 20-series GPU or newer, with at least 8 GB VRAM.</li></ul>Load the workflow, write your first prompt, and give that idea a chance to surprise you.<h2 id="help-artius-moss-reach-more-creators">❤️ Help Artius Moss reach more creators</h2>If you enjoy the model, leave a like, add it to your favorites, or drop a comment. It’s a small gesture that helps more people discover Artius Moss and means a lot to me as its creator.And if you make something you’re happy with, share your results here. I’d love to see where you take it. Feedback about what works well — and what could be better — is welcome too.Thanks for giving Artius Moss a try and helping it grow. ❤️<hr />🔞 18+ only. The model and its custom text encoder are uncensored and support NSFW content. Please follow Civitai’s content rules.Built on MOSS-Image by OpenMOSS (Apache 2.0). Thanks to the original model authors, the LoRA creators, the TSV-Merge researchers, and the ComfyUI team.

Artius Moss

Files with SHA256: b0a6c1ab290b55c083fea294179c6c41b9027f645355e7eb63566dd1cb63efd1