FFF_MOSS-Alba

作者：flyfront

Latest (2026-11-06): v1.1. v1 with a small second fine-tune blended in; nearly the same pictures, a little less blush by default. The matching LoRA is FFF_MOSS-Alba-style v2.1 (https://civitai.com/models/3017759/fffmoss-alba-style). 最新(2026-11-06)は v1.1 です。v1 に短い 2 回目の fine-tune を少し混ぜたもので、絵はほぼ同じまま、既定の赤面が少し減っています。対応する LoRA は FFF_MOSS-Alba-style v2.1 です。

MOSS-Image SFT with the style of FFF_ANIMA-Alba built in: people come out as anime illustrations in that style, with no LoRA and no trigger word. Only the people were trained, so the scenery and text keep MOSS-Image's detail. FFF_ANIMA-Alba の画風を組み込んだ MOSS-Image SFT です。LoRA も trigger も無しで、人物がその画風のアニメ絵になります。学習したのは人物だけなので、背景と文字は MOSS-Image の描き込みを生かしています。 Full card on Hugging Face (prompt, negative prompt, settings, samples, license): https://huggingface.co/flyfront/FFF_MOSS-Alba 詳しい使い方(プロンプト・ネガティブ・設定・生成例・ライセンス)は Hugging Face のページにあります。

Quick start

• Use it like MOSS-Image SFT: Qwen3.5-9B text encoder, FLUX.2 small-decoder VAE. The INT8 ConvRot file (recommended) gives almost the same pictures as bf16 at about half the time.
• Danbooru-style tags or plain English sentences both work. No trigger. Put best quality, highres anywhere in every prompt.
• Euler, 25-30 steps, MOSS or Simple scheduler (shift 5), CFG 3. About 2.4M pixels in any aspect ratio (1344x1792, 1792x1344, 1536x1536); down to 1M pixels also works.
• Negative prompt: the usual MOSS-Image one (full list on the Hugging Face card) plus blush, embarrassed unless you want a blush (SFT blushes by default), overexposure to keep bright daylight scenes from washing out, and warm lighting to keep an orange cast off the whole picture.
• Text in quotes is drawn as written, as with MOSS-Image SFT. Other looks ("realistic photo", "watercolor painting") still respond, but more weakly than in SFT.
• 1boy tends to come out as a girl; describe him in a sentence instead.

About

A full fine-tune of MOSS-Image Base on 1,377 pictures generated with FFF_ANIMA-Alba (tags and natural-language captions, people only in the loss), with the learned weight change added to MOSS-Image SFT. The same style is also available as a LoRA for other MOSS-Image Base-family checkpoints: FFF_MOSS-Alba-style v2 (https://civitai.com/models/3017759/fffmoss-alba-style). FFF_ANIMA-Alba で生成した 1,377 枚(タグと自然文の caption、loss は人物だけ)で MOSS-Image Base を full fine-tune し、変わった重みを MOSS-Image SFT に足しました。同じ画風は、ほかの MOSS-Image Base 系 checkpoint 向けの LoRA(FFF_MOSS-Alba-style v2)としても出しています。

License

Apache License 2.0, the same as the MOSS-Image weights. Generated images are yours. MOSS-Image の重みと同じ Apache License 2.0 です。生成した画像は利用者のものです。