flyfront commited on
Commit
aab9ef9
·
verified ·
1 Parent(s): 4953c6f

Add FFF_MOSS-Alba-v1 (model) and FFF_MOSS-Alba-style-v2 (LoRA)

Browse files
.gitattributes CHANGED
@@ -35,3 +35,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  samples/classroom.png filter=lfs diff=lfs merge=lfs -text
37
  samples/street.png filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  samples/classroom.png filter=lfs diff=lfs merge=lfs -text
37
  samples/street.png filter=lfs diff=lfs merge=lfs -text
38
+ samples/v1_classroom.png filter=lfs diff=lfs merge=lfs -text
39
+ samples/v1_layout.png filter=lfs diff=lfs merge=lfs -text
40
+ samples/v1_layout_boxes.png filter=lfs diff=lfs merge=lfs -text
41
+ samples/v1_street.png filter=lfs diff=lfs merge=lfs -text
42
+ samples/v1_text.png filter=lfs diff=lfs merge=lfs -text
FFF_MOSS-Alba-style-v2.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b41408ac8511afb484bd08f5de88e5c8d3befc2f165d332f1bc13fe41cb9bf33
3
+ size 297355008
FFF_MOSS-Alba-style-v2.safetensors.sha256 ADDED
@@ -0,0 +1 @@
 
 
1
+ b41408ac8511afb484bd08f5de88e5c8d3befc2f165d332f1bc13fe41cb9bf33 FFF_MOSS-Alba-style-v2.safetensors
FFF_MOSS-Alba-v1.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d394b0230b6301310a3b070fe49c9ff42c6824c993a642077f61cac7ad1dca59
3
+ size 10164656328
FFF_MOSS-Alba-v1.safetensors.sha256 ADDED
@@ -0,0 +1 @@
 
 
1
+ d394b0230b6301310a3b070fe49c9ff42c6824c993a642077f61cac7ad1dca59 FFF_MOSS-Alba-v1.safetensors
FFF_MOSS-Alba-v1_int8_convrot.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5c1b91c53cf4a36e44485fb028fc4329f84e567a644a21aa26e40b6649223d40
3
+ size 5413699358
FFF_MOSS-Alba-v1_int8_convrot.safetensors.sha256 ADDED
@@ -0,0 +1 @@
 
 
1
+ 5c1b91c53cf4a36e44485fb028fc4329f84e567a644a21aa26e40b6649223d40 FFF_MOSS-Alba-v1_int8_convrot.safetensors
README.md CHANGED
@@ -1,138 +1,176 @@
1
  ---
2
  license: apache-2.0
3
  base_model:
4
- - OpenMOSS-Team/MOSS-Image-Base
5
- base_model_relation: adapter
6
  pipeline_tag: text-to-image
7
  tags:
8
  - moss-image
9
- - lora
10
  - anime
11
- - style
12
  - text-to-image
 
13
  - safetensors
14
  language:
15
  - en
16
  - ja
17
  ---
18
 
19
- # FFF_MOSS-Alba-style v1
 
 
20
 
21
- A style LoRA for [MOSS-Image](https://github.com/OpenMOSS/MOSS-Image). It puts the look of [FFF_ANIMA-Alba](https://huggingface.co/flyfront/FFF_ANIMA-Alba) (soft shading, warm skin, clear eyes) on the people in the picture and leaves MOSS-Image's backgrounds as they are.
22
 
23
- [MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) 用の画風 LoRA です。人物に [FFF_ANIMA-Alba](https://huggingface.co/flyfront/FFF_ANIMA-Alba) の画風(柔らかい塗り、暖かい肌、はっきりした瞳)を乗せ、背景は MOSS-Image のまま残します。
 
 
 
 
 
 
 
 
 
24
 
25
  ## Samples
26
 
27
  <table>
28
  <tr>
29
- <td align="center" valign="middle" width="50%"><img src="samples/classroom.png" alt="classroom" width="400"></td>
30
- <td align="center" valign="middle" width="50%"><img src="samples/street.png" alt="street" width="400"></td>
 
 
 
 
31
  </tr>
32
  </table>
33
 
34
- Both samples: MOSS-Image SFT + this LoRA at 1.0, tags only, Euler, 25 steps, 1344x1792. Left: CFG 3. Right: CFG 5. Prompts, negative prompts and seeds are in the PNG metadata.
 
 
 
 
 
 
35
 
36
- 生成例は 2 枚とも MOSS-Image SFT + この LoRA(1.0)、タグだけのプロンプト、Euler、25 steps、1344x1792 です。左は CFG 3、右は CFG 5。プロンプト・ネガティブ・seed は PNG のメタデータにあります。
37
 
38
- ## Quick start
 
 
39
 
40
  ```
41
- @fyfnt, <your tags> <lora:FFF_MOSS-Alba-style-v1:1>
42
  ```
43
 
44
  Negative prompt:
45
 
46
  ```
47
- nsfw, worst quality, bad anatomy, watermark, jpeg artifact, blush, embarrassed, overexposure
48
  ```
49
 
50
- Recommended settings / 推奨設定 (the samples above use these at 25 steps; the right one at CFG 5 / 上の生成例は 25 steps、右は CFG 5):
51
-
52
  | | Recommended / 推奨 | Also works / 他の選択肢 |
53
  |---|---|---|
54
- | Checkpoint | MOSS-Image **SFT** | Base: calmer, cooler palette |
55
- | Trigger | `@fyfnt` at the head of the prompt | none: a slightly more settled face |
56
- | LoRA weight | **1.0** | lower weights bring back SFT's own habits |
57
- | Sampler | Euler, **25-30 steps**, MOSS scheduler (shift 5) | Simple scheduler with shift 5 works just as well (the samples use it) |
58
- | CFG | **3** (soft shading) | 5-6 for a cel-shaded look |
59
- | Size | about **2.4M pixels**, any aspect ratio: 1344x1792 (portrait), 1792x1344 (landscape), 1536x1536 (square) | down to about 1M pixels (896x1152, 1024x1024, ...) |
60
- | Quality tags | none | |
 
 
 
 
 
 
61
 
62
- ## How to use it
63
 
64
- **Base or SFT.** The LoRA was trained on MOSS-Image Base and works on both. SFT gives warmer colors, more saturation and richer backgrounds; Base gives a calmer, cooler palette closer to the teacher. The samples use SFT.
65
 
66
- **Trigger is optional.** The style comes through with or without `@fyfnt`. With it, eyes are a little larger and the face is closer to the teacher; without it, the face is a little more settled. Use whichever you prefer. Its position does not matter: the head of the prompt, the end, or next to the LoRA tag all work the same.
67
 
68
- **CFG picks the finish.** CFG 3 gives soft shading with gentle gradients. Around 6 the look turns toward cel shading with stronger lines and more saturation. 7 and above starts to crush dark areas.
69
 
70
- **Blush.** MOSS-Image SFT likes to blush. `blush, embarrassed` in the negative prompt removes it without changing anything else; write `light blush` or `blush` in the prompt when you want it (and drop them from the negative).
71
 
72
- **Bright scenes.** `overexposure` in the negative prompt keeps daylight scenes from washing out; the LoRA's pale skin plus a light outfit and a bright street can otherwise push the whole picture toward white.
73
 
74
- **Sheer fabric.** Lace sleeves and thin skirts tend to show the body through them. Add `see-through clothes` to the negative prompt when you want the fabric opaque.
75
 
76
- **Weight stays at 1.0.** Lowering the weight brings back SFT's own habits (blush, cleavage) rather than softening the style.
77
 
78
- **Color words spread.** MOSS-Image reads tags as context, so `white dress` can also lighten hair, skin and the whole picture. Balance a white outfit with a hair color or a darker room, or let one dark garment anchor the palette. Name the hair color when you want it fixed; otherwise it drifts between black and silver.
79
 
80
- **Distilled LoRAs.** Works together with the 8-step DMD, 4-step and 2-step LoRAs at 1.0 each (8-step is the best trade). At CFG 1 the negative prompt is not evaluated, so use [sd-forge-negpip](https://github.com/flyfront/sd-forge-negpip) (MOSS-Image support) and write `(blush:-1.0), (embarrassed:-1.0)` in the positive prompt instead.
81
 
82
- **Hands.** Fingers can come out simplified, as with other style LoRAs on MOSS-Image. ADetailer-style hand fixing helps. Keep pose words to one per kind (distance, angle, posture) so the model is not asked for an impossible pose.
83
 
84
- **Backgrounds.** Nothing to add: the LoRA was trained with the background masked out, so `detailed background` and similar tags make no difference and MOSS-Image draws the scene the way it always does.
85
 
86
- ## 使い方
87
 
88
- **Base でも SFT でも。** MOSS-Image Base で学習した LoRA ですが、どちらにも効きます。SFT は暖色で彩度が高く背景の描き込みが多め、Base は落ち着いた寒色寄りで教師に近い色味です。生成例は SFT です。
89
 
90
- **trigger は任意。** `@fyfnt` があってもなくても画風は出ます。付けると瞳が少し大きく教師寄りの顔、外すと少し落ち着いた顔立ちになります。位置はどこでも構いません(先頭でも末尾でも LoRA タグの隣でも同じ)。
91
 
92
- **CFG で仕上がりを選ぶ。** 3 は柔らかい塗りで階調が残ります。6 前後でセル調に寄り、線が強く彩度が上がります。7 以上は暗部が潰れ始めます。
93
 
94
- **赤面。** MOSS-Image SFT は赤面しがちです。ネガティブに `blush, embarrassed` を入れると他を変えずに消えます。赤面させたいときはプロンプトに `light blush` か `blush` を書き、ネガから外してください。
95
 
96
- **明るい場面。** ネガティブの `overexposure` は昼の屋外が白く飛ぶのを抑えます。LoRA の淡い肌に明るい服と明るい街が重なると、画面全体が白に寄りがちです。
97
 
98
- **透ける布。** レースの袖や薄いスカートは体が透けがちです。透けさせたくないときはネガティブに `see-through clothes` を足してください。
99
 
100
- **倍率は 1.0 のまま。** 下げると画風が薄まるより先に SFT 自身の癖(赤面、胸元)が戻ります。
101
 
102
- **色の語は全体に広がる。** MOSS-Image はタグを文脈として読むので、`white dress` が髪や肌、画面全体まで白っぽくします。白い服のときは髪色や暗い部屋で釣り合いを取るか、濃い色の服を 1 つ置いてください。髪色は書かないと黒と銀の間で揺れます。固定したいときは書きます。
 
 
 
 
 
 
103
 
104
- **蒸留 LoRA との併用。** 8-step DMD、4-step、2-step の各 LoRA と両方 1.0 で重ねられます(8-step が釣り合いがよい)。CFG 1 ではネガティブが評価されないので、[sd-forge-negpip](https://github.com/flyfront/sd-forge-negpip)(MOSS-Image 対応)で正のプロンプトに `(blush:-1.0), (embarrassed:-1.0)` と書いてください。
 
 
105
 
106
- **手。** MOSS-Image に画風 LoRA を当てたときの常で、指が簡略になることがあります。ADetailer 系の手の補正が効きます。ポーズの語は種類ごと(距離・角度・体勢)に 1 つずつにして、無理な体勢を頼まないのがこつです。
107
 
108
- **背景。** 足す語はありません。背景をマスクして学習しているので、`detailed background` などの有無で絵は変わらず、MOSS-Image がいつもどおり描きます。
 
 
109
 
110
  ## Files
111
 
112
  | File | Size | SHA256 |
113
  |---|---|---|
 
 
 
114
  | `FFF_MOSS-Alba-style-v1.safetensors` | 74.4 MB | `4e54872c45dcb4025777cd5def4dba269739e53a388f881f7b2662b219ca9a08` |
115
 
116
- LoRA on the DiT only (dim 16, alpha 1, all 42 blocks), sd-scripts key layout (`lora_unet_layers_*`). Loads as-is in Forge Neo and ComfyUI; no text encoder weights are included.
117
 
118
- DiT だけの LoRA(dim 16、alpha 1、42 block 全部)、sd-scripts のキー形式(`lora_unet_layers_*`)。Forge Neo と ComfyUI でそのまま読めます。text encoder の重みは含みません。
119
 
120
  ## Training
121
 
122
- Trained with [sd-scripts](https://github.com/kohya-ss/sd-scripts) (`moss_image_train_network.py`) on MOSS-Image Base, with the Qwen3.5-9B text encoder frozen. The teacher set is 1,457 images generated with FFF_ANIMA-Alba, captioned with the tags they were generated from. The loss was masked to the people in each image (background weight 0, hands excluded), so only the character rendering is learned. 1024-pixel buckets, 3 epochs (6,348 steps), LR 1e-4, AdamW8bit, constant schedule.
123
-
124
- [sd-scripts](https://github.com/kohya-ss/sd-scripts)(`moss_image_train_network.py`)で MOSS-Image Base を起点に学習し、Qwen3.5-9B の text encoder は凍結しました。教師は FFF_ANIMA-Alba で生成した 1,457 枚で、caption は生成に使ったタグです。loss は人物の領域だけに掛け(背景の重み 0、手は除外)、キャラクターの描き方だけを学ばせています。1024 画素のバケット、3 epoch(6,348 step)、LR 1e-4、AdamW8bit、constant。
125
 
126
- ## About this repository
127
 
128
- This is the MOSS-Image branch of the FFF_ANIMA-Alba series. A full fine-tune of MOSS-Image on the same teacher may be added here later under the name `FFF_MOSS-Alba`; the LoRA keeps the `-style` suffix.
129
 
130
- FFF_ANIMA-Alba シリーズの MOSS-Image 版です。同じ教師で MOSS-Image を full fine-tune したモデルを、将来 `FFF_MOSS-Alba` の名前でここに足すかもしれません。LoRA は `-style` 付きの名前のままです。
131
 
132
  ## License
133
 
134
- Apache License 2.0 (`LICENSE`), the same as the MOSS-Image weights. The teacher images were generated with FFF_ANIMA-Alba, whose license (CircleStone Labs Non-Commercial License v1.2) places no restriction on the use of generated images, so nothing from that license carries over to this LoRA. Generated images are yours.
135
 
136
- Apache License 2.0(`LICENSE`)。MOSS-Image の重みと同じです。教師画像は FFF_ANIMA-Alba で生成したもので、その license(CircleStone Labs Non-Commercial License v1.2)は生成画像の利用を制限していないため、この LoRA には引き継がれません。生成した画像は利用者のものです。
137
 
138
- Civitai page: https://civitai.com/models/3017759/fffmoss-alba-style
 
1
  ---
2
  license: apache-2.0
3
  base_model:
4
+ - OpenMOSS-Team/MOSS-Image-SFT
5
+ base_model_relation: finetune
6
  pipeline_tag: text-to-image
7
  tags:
8
  - moss-image
 
9
  - anime
10
+ - illustration
11
  - text-to-image
12
+ - lora
13
  - safetensors
14
  language:
15
  - en
16
  - ja
17
  ---
18
 
19
+ # FFF_MOSS-Alba
20
+
21
+ [MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) that draws people in the style of [FFF_ANIMA-Alba](https://huggingface.co/flyfront/FFF_ANIMA-Alba). Only the people were trained (backgrounds were masked out), so the scenery and text keep MOSS-Image's detail.
22
 
23
+ 人物を [FFF_ANIMA-Alba](https://huggingface.co/flyfront/FFF_ANIMA-Alba) の画風で描くようにした [MOSS-Image](https://github.com/OpenMOSS/MOSS-Image) です。学習したのは人物だけ(背景はマスクで除外)なので、背景と文字は MOSS-Image の描き込みを生かしています。
24
 
25
+ > **Latest (2026-11):** the fine-tuned model **FFF_MOSS-Alba-v1** and its LoRA **FFF_MOSS-Alba-style-v2**. FFF_MOSS-Alba-style-v1 is the earlier LoRA, kept for reference.
26
+ >
27
+ > 最新(2026-11)は、fine-tune したモデル FFF_MOSS-Alba-v1 と、その LoRA の FFF_MOSS-Alba-style-v2 です。FFF_MOSS-Alba-style-v1 は以前の LoRA で、参考として残しています。
28
+
29
+ | File | What it is | Use it with |
30
+ |---|---|---|
31
+ | `FFF_MOSS-Alba-v1_int8_convrot.safetensors` | Latest: the model (INT8 ConvRot, recommended) | on its own |
32
+ | `FFF_MOSS-Alba-v1.safetensors` | Latest: the model (bf16) | on its own |
33
+ | `FFF_MOSS-Alba-style-v2.safetensors` | Latest: the same style as a LoRA | other MOSS-Image Base-family checkpoints |
34
+ | `FFF_MOSS-Alba-style-v1.safetensors` | Earlier style LoRA (2026-10) | MOSS-Image Base / SFT |
35
 
36
  ## Samples
37
 
38
  <table>
39
  <tr>
40
+ <td align="center" valign="top" width="50%"><img src="samples/v1_classroom.png" alt="classroom" width="400"><br>Tags, CFG 3</td>
41
+ <td align="center" valign="top" width="50%"><img src="samples/v1_street.png" alt="street" width="400"><br>Tags, CFG 5</td>
42
+ </tr>
43
+ <tr>
44
+ <td align="center" valign="top" width="50%"><img src="samples/v1_text.png" alt="text" width="400"><br>A sentence with quoted text, CFG 3</td>
45
+ <td align="center" valign="top" width="50%"><img src="samples/v1_layout.png" alt="layout" width="400"><br>A JSON layout prompt, CFG 3</td>
46
  </tr>
47
  </table>
48
 
49
+ <p align="center"><img src="samples/v1_layout_boxes.png" alt="the boxes of the layout sample in Regional Boxes" width="300"><br>The boxes behind the layout sample: 1 title, 2 girl, 3 cat (Regional Boxes in Forge Neo)<br>レイアウトの生成例の箱: 1 題字、2 少女、3 猫(Forge Neo の Regional Boxes)</p>
50
+
51
+ All four: FFF_MOSS-Alba-v1 (INT8 ConvRot), Euler, Simple scheduler (shift 5), 25 steps, 1344x1792. Prompts, negative prompts and seeds are in the PNG metadata. The top two have the same subjects as the style-v1 samples.
52
+
53
+ 4 枚とも FFF_MOSS-Alba-v1(INT8 ConvRot)、Euler、Simple(shift 5)、25 steps、1344x1792 です。上の 2 枚はタグ、左下は文字を引用符で書いた文、右下は JSON のレイアウト指定です。プロンプト・ネガティブ・seed は PNG のメタデータにあります。上の 2 枚は style-v1 の生成例と同じ題材です。
54
+
55
+ ## FFF_MOSS-Alba-v1 (the model)
56
 
57
+ MOSS-Image SFT with the style built in. No LoRA, no trigger word. Use it like MOSS-Image SFT: same text encoder (Qwen3.5-9B), same VAE (FLUX.2 small decoder), same settings.
58
 
59
+ スタイルを組み込んだ MOSS-Image SFT です。LoRA も trigger も要りません。text encoder(Qwen3.5-9B)、VAE(FLUX.2 small decoder)、設定は MOSS-Image SFT と同じです。
60
+
61
+ ### Quick start
62
 
63
  ```
64
+ best quality, highres, <your tags or sentences>
65
  ```
66
 
67
  Negative prompt:
68
 
69
  ```
70
+ nsfw, worst quality, bad anatomy, watermark, jpeg artifact, blush, embarrassed, overexposure, warm lighting
71
  ```
72
 
 
 
73
  | | Recommended / 推奨 | Also works / 他の選択肢 |
74
  |---|---|---|
75
+ | File | `FFF_MOSS-Alba-v1_int8_convrot` | bf16: almost the same pictures, about twice the time |
76
+ | Prompt | Danbooru-style tags or plain English sentences | both were trained |
77
+ | Quality tags | `best quality, highres` in every prompt, anywhere (the samples have it at the end) | |
78
+ | Trigger | none | |
79
+ | Sampler | Euler, 25-30 steps, MOSS scheduler (shift 5) | Simple scheduler with shift 5 |
80
+ | CFG | 3 | |
81
+ | Size | about 2.4M pixels, any aspect ratio: 1344x1792, 1792x1344, 1536x1536 | down to about 1M pixels (896x1152, 1024x1024, ...) |
82
+
83
+ ### How to use it
84
+
85
+ Tags or sentences. The model was trained with both a tag caption and a natural-language caption for every picture, so `1girl, solo, black hair, school uniform, classroom` and "A girl with long black hair in a school uniform stands by a classroom window" both work.
86
+
87
+ Anime by default. People come out in the Alba style unless you ask for something else, even in sentences that MOSS-Image SFT draws as photos. Other looks still respond, but more weakly than in SFT: "realistic photo" and "food photography" stay realistic but come out smoother and more polished than a real photo, and "watercolor painting" gives a lighter watercolor touch.
88
 
89
+ Text. MOSS-Image draws the text you quote in the prompt (`a chalkboard sign that reads "FFF MOSS Alba"`, the bottom-left sample), and this model keeps that: English and Chinese signs, posters and mugs came out as written in our tests.
90
 
91
+ Layouts. MOSS-Image's JSON layout prompts (a `summary` and a list of items, each with a `bbox`; see the [MOSS-Image user guide](https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md#layout-prompts-experimental)) work too. In our tests the title, the girl and the cat went where their boxes put them. Give each item a box it fits in: a wide frilled skirt spilled over a box that was too narrow. In Forge Neo, the [Regional Boxes](https://github.com/eduardoabreu81/sd-regional-boxes-neo) extension writes the JSON from boxes you draw on the canvas.
92
 
93
+ Blush. MOSS-Image SFT blushes easily, and so does this model: without `blush, embarrassed` in the negative prompt, about one picture in five shows a clear blush. With them, none did in our tests. When you want one, drop them from the negative and write `blush` or `light blush` in the prompt: in our tests both gave a soft pink on the cheeks, about equally strong.
94
 
95
+ Bright scenes. `overexposure` in the negative prompt keeps daylight scenes from washing out.
96
 
97
+ Warm cast. Without `warm lighting` in the negative prompt, an orange light tends to spread over the whole picture, backgrounds included. With it, whites stay white and skin stays healthy (not pale), while warmth that belongs to the scene (a sunset, a fireplace) is kept.
98
 
99
+ Men. The tag `1boy` tends to come out as a girl, because the training pictures are mostly girls. Describing him in a sentence ("a young man in a dark coat ...") works.
100
 
101
+ Speed. INT8 ConvRot runs at about 11 s per picture against about 24 s for bf16 (896x1152, 30 steps, RTX 4090, Forge Neo) and gives almost the same picture. It has the same layout as Comfy-Org's `moss_image_*_int8_convrot` files.
102
 
103
+ ### 使い方
104
 
105
+ タグでも文でも。学習では 1 枚ごとにタグの caption と自然文の caption の両方を使ったので、`1girl, solo, black hair, school uniform, classroom` も「A girl with long black hair in a school uniform stands by a classroom window」も使えます。
106
 
107
+ 既定はアニメ絵です。何も書かなければ人物は Alba の画風になります(MOSS-Image SFT なら写真で描くような文でも)。ほかの絵柄の指定も効きますが、SFT より弱まります。「realistic photo」「food photography」は写実のままですが、本物の写真より滑らかで整った仕上がりになり、「watercolor painting」は水彩の風合いが薄くなります。
108
 
109
+ 文字。MOSS-Image はプロンプトで引用した文字(`a chalkboard sign that reads "FFF MOSS Alba"`、左下の生成例)を描けます。このモデルもそのままで、英語と中国語の看板・ポスター・マグカップは試した範囲で書いたとおりに出ました。
110
 
111
+ レイアウト。MOSS-Image の JSON のレイアウト指定(`summary` と、`bbox` を持つ項目の並び。書き方は [MOSS-Image のユーザーガイド](https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md#layout-prompts-experimental))も使えます。試した範囲では、題字・少女・猫は箱で指定した位置に置かれました。箱は中身が収まる大きさにしてください。広がったフリルのスカートは、狭すぎる箱からはみ出しました。Forge Neo では、拡張の [Regional Boxes](https://github.com/eduardoabreu81/sd-regional-boxes-neo) を使うと、画面に描いた箱から JSON を書いてくれます。
112
 
113
+ 赤面。MOSS-Image SFT は赤面しがちで、このモデルも同じです。ネガティブに `blush, embarrassed` が無いと 5 枚に 1 枚ほどはっきり赤面し、入れると試した範囲では出ませんでした。赤面させたいときは、ネガから外してプロンプトに `blush` か `light blush` を書いてください。試した範囲では、どちらも頬がほんのり色づく程度で、強さはほぼ同じでした。
114
 
115
+ 明るい場面。ネガティブの `overexposure` は昼の屋外が白く飛ぶのを抑えます。
116
 
117
+ 暖色のかぶり。ネガティブに `warm lighting` が無いと、背景も含めて画面全体に橙色の光がかかりがちです。入れると白は白のまま、肌は健康的なまま(蒼白にならず)で、夕焼けや暖炉のように場面に理由がある暖かさは残ります。
118
 
119
+ 男性。教師がほとんど少女なので、タグの `1boy` は少女になりがちです。文で書けば(「a young man in a dark coat ...」)男性になります。
120
 
121
+ 速度。INT8 ConvRot は 1 枚約 11 秒、bf16 は約 24 秒(896x1152、30 steps、RTX 4090、Forge Neo)で、絵はほぼ同じです。Comfy-Org の `moss_image_*_int8_convrot` と同じ形式です。
122
 
123
+ ## FFF_MOSS-Alba-style-v2 (LoRA)
124
 
125
+ The style of FFF_MOSS-Alba-v1 as a LoRA (rank 64), for other checkpoints built on MOSS-Image Base. On MOSS-Image SFT itself, use FFF_MOSS-Alba-v1 instead.
126
 
127
+ FFF_MOSS-Alba-v1 の画風を LoRA(rank 64)にしたもので、MOSS-Image Base を元にしたほかの checkpoint 向けです。MOSS-Image SFT で使うなら、FFF_MOSS-Alba-v1 を使ってください。
128
 
129
+ ```
130
+ best quality, highres, <your tags or sentences> <lora:FFF_MOSS-Alba-style-v2:1>
131
+ ```
132
+
133
+ - No trigger word. Keep the weight at 1.0. Same negative prompt as above (`blush, embarrassed` matters more here: without it, blush comes out more often than with FFF_MOSS-Alba-v1).
134
+ - Use it on an INT8 or FP8 checkpoint. The LoRA is the fine-tune's weight change, which is about the size of one bf16 rounding step; Forge Neo and ComfyUI write a LoRA into bf16 weights with round-to-nearest, which drops part of it and changes the picture, while INT8 / FP8 weights are re-quantized with stochastic rounding and keep it on average.
135
+ - Tested on [Kromoss](https://huggingface.co/lodestones/Kromoss) (INT8 ConvRot) and MOSS Cat Tower (INT8): the style comes through on both. How often a blush appears depends on the checkpoint (Cat Tower blushes a lot on its own), so keep `blush, embarrassed` in the negative prompt.
136
 
137
+ - trigger はありません。倍率は 1.0 のままで。ネガティブは上と同じです(`blush, embarrassed` はこちらの方が大事で、無いと FFF_MOSS-Alba-v1 より赤面が多く出ます)。
138
+ - INT8 か FP8 の checkpoint に当ててください。この LoRA は fine-tune で変わった重みの差で、大きさが bf16 の丸めの刻みほどしかありません。Forge Neo と ComfyUI は bf16 の重みには LoRA を最も近い値への丸めで書き込むので一部が消えて絵が変わり、INT8 / FP8 の重みは確率的な丸めで量子化し直すので平均では残ります。
139
+ - [Kromoss](https://huggingface.co/lodestones/Kromoss)(INT8 ConvRot)と MOSS Cat Tower(INT8)で試し、どちらにも画風が乗りました。赤面の出やすさは checkpoint によって違う(Cat Tower は元から赤面が多い)ので、ネガティブの `blush, embarrassed` は残してください。
140
 
141
+ ## FFF_MOSS-Alba-style-v1 (earlier LoRA, superseded)
142
 
143
+ The first release (2026-10-24): a style LoRA (dim 16) trained directly on MOSS-Image Base with the background masked out. Use it on MOSS-Image SFT (or Base) at 1.0 with `@fyfnt` at the head of the prompt (optional) and the same negative prompt. Its samples are `samples/classroom.png` and `samples/street.png`. v2 and the model above come instead from a full fine-tune on the same kind of teacher pictures, with natural-language captions as well as tags.
144
+
145
+ 最初の公開(2026-10-24)。MOSS-Image Base に背景をマスクして直接学習した画風 LoRA(dim 16)です。MOSS-Image SFT(または Base)に倍率 1.0 で当て、プロンプトの先頭に `@fyfnt`(任意)、ネガティブは上と同じです。生成例は `samples/classroom.png` と `samples/street.png`。v2 と上のモデルは、同じ種類の教師画像で full fine-tune したもので、caption にタグだけでなく自然文も使っています。
146
 
147
  ## Files
148
 
149
  | File | Size | SHA256 |
150
  |---|---|---|
151
+ | `FFF_MOSS-Alba-v1_int8_convrot.safetensors` | 5.41 GB | `5c1b91c53cf4a36e44485fb028fc4329f84e567a644a21aa26e40b6649223d40` |
152
+ | `FFF_MOSS-Alba-v1.safetensors` | 10.16 GB | `d394b0230b6301310a3b070fe49c9ff42c6824c993a642077f61cac7ad1dca59` |
153
+ | `FFF_MOSS-Alba-style-v2.safetensors` | 297 MB | `b41408ac8511afb484bd08f5de88e5c8d3befc2f165d332f1bc13fe41cb9bf33` |
154
  | `FFF_MOSS-Alba-style-v1.safetensors` | 74.4 MB | `4e54872c45dcb4025777cd5def4dba269739e53a388f881f7b2662b219ca9a08` |
155
 
156
+ The model files are DiT only, in the Comfy-Org single-file layout (fused `proj.qkv`); they load in Forge Neo and ComfyUI like `moss_image_sft_*`. The LoRAs use the sd-scripts key layout (`lora_unet_layers_*`, separate q / k / v). No text encoder weights are included.
157
 
158
+ モデルのファイルは DiT だけで、Comfy-Org の単一ファイル形式(`proj.qkv` をまとめた形)です。Forge Neo と ComfyUI で `moss_image_sft_*` と同じように読めます。LoRA は sd-scripts のキー形式(`lora_unet_layers_*`、q / k / v 別)です。text encoder の重みは含みません。
159
 
160
  ## Training
161
 
162
+ A full fine-tune of MOSS-Image Base with [Fizgig](https://github.com/shootthesound/Fizgig): 1,377 pictures generated with FFF_ANIMA-Alba, each with two captions (the tags it was generated from, and a plain-English description written by Qwen3-VL-4B with the tags as hints). The loss was masked to the people (background weight 0), so MOSS-Image keeps drawing the scenes. 1 MP, LR 1e-5, Adafactor, 2 rotations (5,508 steps). The text encoder was not trained.
 
 
163
 
164
+ FFF_MOSS-Alba-v1 is MOSS-Image SFT plus the full weight change of that fine-tune (fine-tuned Base minus Base). FFF_MOSS-Alba-style-v2 is the same change reduced to rank 64 by SVD. The INT8 ConvRot file was made with `quantize_convrot.py` from [lodestones/Kromoss](https://huggingface.co/lodestones/Kromoss), which reproduces Comfy-Org's MOSS-Image INT8 files bit for bit.
165
 
166
+ [Fizgig](https://github.com/shootthesound/Fizgig) で MOSS-Image Base を full fine-tune しました。教師は FFF_ANIMA-Alba で生成した 1,377 枚で、1 枚ごとに caption が 2 つ(生成に使ったタグと、タグを手がかりに Qwen3-VL-4B が書いた英語の説明文)あります。loss は人物の領域だけに掛け(背景の重み 0)、背景は MOSS-Image の描き方を残しています。1MP、LR 1e-5、Adafactor、2 rotation(5,508 step)。text encoder は学習していません。
167
 
168
+ FFF_MOSS-Alba-v1 は、この fine-tune で変わった重み(fine-tune 後の Base − Base)を全部 MOSS-Image SFT に足したものです。FFF_MOSS-Alba-style-v2 は同じ差を SVD で rank 64 にしたものです。INT8 ConvRot は [lodestones/Kromoss](https://huggingface.co/lodestones/Kromoss) の `quantize_convrot.py` で作りました(Comfy-Org の MOSS-Image の INT8 ファイルをビット単位で再現する作りです)。
169
 
170
  ## License
171
 
172
+ Apache License 2.0 (`LICENSE`), the same as the MOSS-Image weights. The teacher pictures were generated with FFF_ANIMA-Alba, whose license (CircleStone Labs Non-Commercial License v1.2) places no restriction on the use of generated images, so nothing from that license carries over here. Generated images are yours.
173
 
174
+ Apache License 2.0(`LICENSE`)。MOSS-Image の重みと同じです。教師画像は FFF_ANIMA-Alba で生成したもので、その license(CircleStone Labs Non-Commercial License v1.2)は生成画像の利用を制限していないため、ここには引き継がれません。生成した画像は利用者のものです。
175
 
176
+ Civitai: style LoRA https://civitai.com/models/3017759/fffmoss-alba-style (the model page is coming soon)
samples/v1_classroom.png ADDED

Git LFS Details

  • SHA256: 6c8260acb975a22df36531091424d54cfc5b7abe0536ae2bb211fc7a3aaff6ef
  • Pointer size: 132 Bytes
  • Size of remote file: 2.76 MB
samples/v1_layout.png ADDED

Git LFS Details

  • SHA256: 99d6081e10e05c033236d24709aea537a2a46367096844ce8ef79c93c2feff5e
  • Pointer size: 132 Bytes
  • Size of remote file: 3.14 MB
samples/v1_layout_boxes.png ADDED

Git LFS Details

  • SHA256: 4462e9e2bce8487cec8eb94ec17ca258c6aedcffe7f82c7954e728b482211f9f
  • Pointer size: 131 Bytes
  • Size of remote file: 767 kB
samples/v1_street.png ADDED

Git LFS Details

  • SHA256: 896b966ec1c6b645d1d1f9033537e3d68648006252e5cee8bbc9f25aec095ff6
  • Pointer size: 132 Bytes
  • Size of remote file: 3.77 MB
samples/v1_text.png ADDED

Git LFS Details

  • SHA256: 15f8d2a25fde269e01117f02a2bfe191bb15211a0642088aa6dd4baa23c4e12f
  • Pointer size: 132 Bytes
  • Size of remote file: 3.16 MB