TIN_Moss_image_sft_int8_8steps Realism (真实感人像)這個模型適合一般使用到偏逼真人像創造的空間使用3060 12GB顯示卡測試在 ComfyUI 和 WebUI Forge Neo，生成速度差不多，用int8模型生成1024x1536，第一次8步生成用16秒左右，第二次8步生成用10秒左右，第一次4步生成用10秒左右，第二次4步生成用6秒左右，(使用 ComfyUI 生成的質量感覺會比 Web ForUIge Neoge要留意在 WebUI Forge Neo 使用要詳細的提示詞，使用較簡單的提示詞時會產生色塊圖或全黑圖,最好的是在 webui-user.bat 文件裡修改 set COMMANDLINE_ARGS=--disable-sage這樣就可以避免生成色塊圖或全黑圖在提示詞加入 HDR photo of cinematic film still. 可以產生更有真實感人像採樣方法 Sampler：Euler迭代步數 Steps：8 或更多 （4步也可以）提示詞引導係數CFG：1圖片尺寸：768 x 1152 &amp; 1024 x 1536使用 qwen3.5_9b_w4a8_gptq.safetensors 會比使用 qwen3.5_9b_int8_convrot.safetensors 產生速度更快一些使用 flux2-vae_small.safetensors 和使用 flux2-vae.safetensors 感覺差不多效果(202601025 直接分享原模型下載)(歡迎大家測試和返圖)mypikpak.com/s/VOcdFHxoJjvcCAewVVS40pQKo2 圖集

TIN_Moss_image_sft_int8_8steps Realism

TIN_Moss_image_sft_int8_8steps Realism

flux2-vae_small.safetensors

tinMossImageSftInt8_int88stepsRealism.safetensors

int8_8steps_Realism

<h2 id="i.-introduction">I. Introduction</h2>MossAD-Demo (Moss Anime Diffusion) is a Danbooru/Gelbooru/E621 proof of concept Full-Finetune of OpenMOSS's MOSS-Image Base 5B T2I model, trained on a small curated 3,685 images dataset of 57 artists, as a way to test how fast and how well MOSS-Image learns artist styles, booru tags prompting and layout (bounding box) prompting before committing to a much bigger run.The test is a big ass success: after only ~39 hours on a single AMD MI300X (ok I know it's alot for just 3k image on this kind of gpu, but i received the compute very suddenly under a 4 day deadline and didn't have the luxury to optimize training), the styles land at around 90% accuracy, they carry over to subjects the artist never drew (a furry artist's style on anime girls, and the other way around), and hands are solid enough barring some exceptions here and there.~ What the pilot found, and what the next big run will use:<ul><li>MOSS-Image learns styles fast: ~4.5k views per artist at 512px already gives 70-75% style accuracy, ~7k views gives 85-90% with the style carrying over to other subjects, while ~4k views only shows the style in the artist's own domain (furry artist styles only on furry gens)</li><li>Most of the training can stay at cheap 512px, with shorter 768px and 1024px stages for hands, details and texture</li><li>The learning rate needs to be halved at every higher resolution stage (4e-5 → 2e-5 → 1e-5), carrying the 512px LR into 768px made it unstable within 5 epochs and the stage had to be restarted</li><li>The base model already understands panels and boxes from ComfyUI's default bbox node, and the finetune keeps that working with the trained styles</li><li>Watermarks REALLY stick: if most of an artist's images share the same watermark, no tagging, captioning or negative prompt removes it. I will test masked training in the future to keep watermarks and signatures out of the loss if deemed viable and not too much of a hassle.</li></ul>Please scroll down to the generation recommendation section for the settings and the artist trigger.Any donation will help with funding the full run, which will begin in a few months time when the OpenMOSS devs release their official anime tune, and I finish my 6k artist style dataset, you can do so through:Crypto (Preferred, cause Kofi/Paypal money transfer time is ass and they take a big cut)0x6a4bc748cd0bb9ced9a360eb0eb79f4f106614f8 (USDT - BEP20 Network - Custodial wallet)12PPVYUeS1MerNp38Tpns5qXR6cmhu9tws (Bitcoin - BTC Network - Custodial wallet)0x6a4bc748cd0bb9ced9a360eb0eb79f4f106614f8 (Ethereum - ERC20 Network - Custodial wallet)GKjLvD9XGAhyEpsZvEUNH5je9PFbrivoV5yBfYKGYskq (Solana - SOL network - Non Custodial wallet)Please handle your money carefully and make sure the address you're sending to is correct. Please contact me through discord (user: talannnn) if you're donating with crypto, to make sure it goes through and you're not losing money.Ko-fi<a target="_blank" rel="ugc" href="https://ko-fi.com/talanartvn">https://ko-fi.com/talanartvn</a><h2 id="ii.-model-details">II. Model Details</h2><ul><li>Base model: <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base">OpenMOSS's MOSS-Image Base 5B</a></li><li>Method: Full Finetune</li><li>Trainer: <a target="_blank" rel="ugc" href="https://github.com/RicemanT/mage-flow-trainer">My mage-flow-trainer fork</a> (of <a target="_blank" rel="ugc" href="https://github.com/bluvoll/mage-flow-trainer">Bluvoll's mage-flow-trainer</a>)</li><li>Hardware: 1x AMD MI300X 192GB (DigitalOcean AMD Developer Cloud), free compute from Disty</li><li>Total training time: ~39 hours of MI300X for the final model (~55 hours including LR tests and the restarted stage)</li><li>Total samples seen: ~457k (~8k views per artist)</li><li>Training resolutions: 512² → 768² → 1024², mixed aspect ratios packed with varlen attention</li></ul><h3 id="training-run">Training run</h3>Budget: 0 dollars (MI300X, free compute baby, but like it prob took 90-100 dollars in reality...thank god i can fuck around for free)Full configs: <a target="_blank" rel="ugc" href="https://huggingface.co/RicemanT/MossAD-Demo/tree/main/configs">configs folder</a><ul><li>512px: peak LR 4e-5, 70 epochs, 560 steps, effective batch 512 (128 x 4), ~4,550 views per artist, ~12.8h</li><li>768px: peak LR 2e-5, 40 epochs, 600 steps, effective batch 256 (64 x 4), ~2,600 views per artist, ~15.5h</li><li>1024px: peak LR 1e-5, 14 epochs, 406 steps, effective batch 128 (32 x 4), ~910 views per artist, ~10.5h</li></ul><ul><li>LR scheduler: Warmup (5%) -&gt; Constant (55%) -&gt; REX to 0 (40%), for every stage</li><li>Precision: Full BF16</li><li>Optimizer: AdamW8bit (quantized state, no Kahan summation)</li><li>Attention: torch varlen with packed resolutions (no caption padding, no bucketing tricks)</li><li>LR test: 4e-5 vs 2e-5 at 512px for 10 epochs each, 4e-5 learned styles a lot faster and was continued</li></ul><h3 id="additional-training-features">Additional training features</h3><ul><li>16 caption variations per image</li><li>Mixed captions at 25/10/15/15/25/20 ratio (tags only, short NL, medium NL, long NL, layout with bounding boxes, tags-nl)</li><li>Layout captions: characters, objects, text, panels and character detail boxes, with short box captions in 30% of the layouts</li><li>Tag dropout: 10%</li><li>Caption dropout: 5%</li><li>Tag shuffle</li><li>Caption shuffle</li><li>Artist trigger attribution system (the trigger always stays at the front)</li></ul><h2 id="iii.-recommended-settings">III. Recommended Settings</h2>The model works in ComfyUI with its native MOSS-Image support: load the checkpoint as the diffusion model, together with the text encoder and VAE from <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image">Comfy-Org/MOSS-Image</a>. Also visit <a target="_blank" rel="ugc" href="https://huggingface.co/kdcth/MossAD-int8-convrot">https://huggingface.co/kdcth/MossAD-int8-convrot</a> for the int8 convrot version of the model. These are the settings used for the sample images above (ComfyUI, all showcase images above have metadata):<ul><li>Shift: 5.0 (or comfyui default)</li><li>Steps: 30</li><li>CFG: 5</li><li>Sampler: euler</li><li>Scheduler: simple</li><li>Negative prompt: worst quality, bad anatomy, watermark, jpeg artifact</li></ul>These are just my usual settings and workflow — feel free to experiment.The model also includes quality tags, you usually want to add them at the end of the tag section of the prompt, they are:masterpiece, best quality, very aesthetic, aestheticArtist Trigger: This model uses the @artist name trigger, if you want to use the model's built-in artist styles, please always put one at the start of your prompt (for example "@sakimichan, 1girl, ..." or "@sakimichan. A girl standing in the rain..."). The model supports 57 artists, please download to see: <a target="_blank" rel="ugc" href="https://huggingface.co/RicemanT/MossAD-Demo/blob/main/MossAD-Demo-artists.csv">MossAD-Demo-artists.csv</a>. The characters the dataset covers are in <a target="_blank" rel="ugc" href="https://huggingface.co/RicemanT/MossAD-Demo/blob/main/MossAD-Demo-characters.csv">MossAD-Demo-characters.csv</a>.All three prompt styles work: booru tags, natural language, and MOSS layout prompts with bounding boxes (ComfyUI's default bbox node on Nodes 2.0 works).<h3 id="known-issues">Known issues</h3><ul><li>Watermarks and signatures that most of an artist's images share can show up in gens</li><li>The model is still a bit unstable and have slightly degraded text performance in general compared to base model</li><li>Hands are kinda bad in some exceptions here and there</li></ul><h2 id="iv.-dataset">IV. Dataset</h2>3,685 images of 57 artists from Danbooru, Gelbooru and E621 (35-75 images per artist), collected and curated with my Illustration Scrapping Studio app, tagged with animetimm Convnext 2 for Danbooru/Gelbooru + RedRocket Hydra 3.5 for E621, and captioned with Gemini 3.8 Flash and Muse Spark 1.3 Contributor, with three lengths of natural language captions and MOSS layout captions, with detector-assisted bounding boxes for text, watermarks, faces and panels.<h2 id="v.-notes-from-the-training-diary">V. Notes from the Training Diary</h2>Full pilot notes (stages, LR lessons, views per artist, ROCm setup): <a target="_blank" rel="ugc" href="https://github.com/RicemanT/mage-flow-trainer/blob/main/scripts/mossad-demo/README.md">MossAD-Demo pilot README</a><h2 id="vi.-license">VI. License</h2>This model is a Derivative of OpenMOSS's MOSS-Image and is distributed under the same <a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0">Apache 2.0 License</a> as the base model, with no additional restrictions.<h2 id="vii.-acknowledgments">VII. Acknowledgments</h2>Beeg thanks to:<ul><li>Disty — The free MI300X compute that made this whole test possible, the biggest thanks to them</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team">OpenMOSS</a> — Creators of MOSS-Image, absolute fuckin legends for providing open source with this amazing base model</li><li><a target="_blank" rel="ugc" href="https://github.com/bluvoll/mage-flow-trainer">Bluvoll</a> — Creator of mage-flow-trainer that my fork is derived from, and general training advice</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/user/motimalu">Motimalu</a> — Inspiration behind finetuning practices and configs, like lowering the LR at every resolution stage (should be obvious but I didn't remember first time)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf">KBlueLeaf</a> — Future sponsor for MossAD</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Anzhc">Anzhc</a> — Detection models used for the bounding box captions, and general training advice</li><li>Format inspired by <a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/blob/main/diffusion-pipe/anima/notes/kirazuri4.0-notes%20.md">Motimalu's Kirazuri diary</a></li></ul>

MossAD Demo - MOSS Image Danbooru/E621 57 Artist Style Finetune

mossadDemoMOSSImage_demo_int8.safetensors

mossadDemoMOSSImage_demo_bf16.safetensors

flux2-vae.safetensors

Demo

<h3 id="moss-image-2.5d-merged-model">🧩 MOSS-Image 2.5D · Merged Model</h3>This is an experimental merge made to test model merging with MOSS-Image. Its main feature is the ability to generate anime images with a 2.5D look directly. The style LoRA is already merged into the model—feel free to try it out and share your results! ✨📜 Attribution, License &amp; DisclaimerThis is an independently modified personal merge, not an official OpenMOSS release or an OpenMOSS-endorsed model. The original MOSS-Image-SFT model is licensed under <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/blob/main/LICENSE">Apache License 2.0</a>. This description does not replace or modify the applicable upstream licenses.Please follow applicable laws and platform rules, and respect others’ copyrights, privacy, likeness rights, and other legal rights. Do not use the model for unlawful activities, infringement, fraud, or other activities that harm others. Before using, publishing, or commercially using any outputs, review them and obtain any necessary permissions. You are responsible for your own use and distribution of the outputs.📝 Prompt SupportThis model supports tag-based prompts, natural-language descriptions, and JSON-formatted prompts, so feel free to use whichever you prefer.Check out the example images for practical references ✨ The original images include ComfyUI workflow and prompt metadata. Download an original image and drag it into ComfyUI to explore the workflow, prompts, and generation settings.⚙️ Generation Tips<ul><li>Recommended generation resolution: 1024 × 1536.</li><li>In my testing so far, anatomy has been reasonably consistent at this resolution, though occasional mistakes still happen.</li><li>Generating directly at 2K or higher resolutions noticeably increases the chance of anatomy and proportion issues. I suggest starting with the recommended resolution.</li></ul>🧪 Model Sources &amp; Merge Recipe<ul><li>Original model: <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">OpenMOSS-Team / MOSS-Image-SFT</a>.</li><li>Merged with: My own MOSS 2.5D style LoRA.</li></ul>Drag the model file into ComfyUI to view the merging workflow and recipe.--------中文说明--------<h3 id="moss-image-2.5d">🧩 MOSS-Image 2.5D · 融合底模</h3>这是一个用于测试 MOSS-Image 模型融合效果的实验性底模，主要特点是能够直接生成带有 2.5D 质感的动漫图片。画风 LoRA 已合入模型，欢迎试用，也欢迎分享你的出图效果！✨📜 来源、许可与免责声明本模型为个人制作的独立融合版本，并非 OpenMOSS 官方发布或背书。原始 MOSS-Image-SFT 模型适用 <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/blob/main/LICENSE">Apache License 2.0</a>，本说明不替代或修改适用的原有许可。请遵守适用法律法规及平台规则，尊重他人的著作权、肖像权、隐私等合法权益，请勿用于违法、侵权、诈骗或其他损害他人权益的活动。使用、发布或商业使用生成内容前，请自行审核并确认必要的授权，对自己的使用和传播行为承担相应责任。📝 提示词支持本模型支持 Tag 标签、自然语言描述，以及 JSON 格式提示词，按自己习惯的方式使用就好。具体写法可以参考我提供的例图 ✨ 示例原图中保留了 ComfyUI 工作流和提示词信息，下载原图后拖入 ComfyUI，就可以查看对应的工作流、提示词和生成参数，方便参考和调整。⚙️ 出图建议<ul><li>推荐生成分辨率：1024 × 1536。</li><li>在我目前的测试中，这个分辨率下的肢体表现比较稳定，但不保证每张图片都没有问题。</li><li>直接生成 2K 或更高分辨率时，肢体和比例出错的概率明显增加，建议先从推荐分辨率开始尝试。</li><li></li></ul>🧪 模型来源与融合配方<ul><li>原始模型： <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">OpenMOSS-Team / MOSS-Image-SFT</a>。</li><li>融合内容： 我自己训练的 MOSS 2.5D 画风 LoRA。</li></ul>将模型文件拖入 ComfyUI，可以查看对应的融合工作流和配方。

Files with SHA256: d8d52ba036475f5fb07c8b435e176d3d97ebfa82f0d1a1c317f9cc1e25bd013b