Transparent-background LoRA for MOSS-Image

A LoRA for MOSS-Image-SFT that makes pictures with an alpha channel. It draws cut-out subjects, logos, lettering and effects on a transparent background, and the sampler outputs RGBA PNGs directly, with no background-removal step. It is used together with trmz's RGBA FLUX.2 VAE (trmz/flux2-klein-alpha), which decodes four channels.

Seed-a outputs over a checker, from the evaluation below. Top row: an anime character, a photo of a fox, a glass of sparkling water, smoke. Bottom row: a tea-shop logo from a JSON layout prompt, brush calligraphy, a roll cake, neon lettering. RGBA originals: examples/*.png.

Files

How to use it

It needs MOSS-Image-SFT (f972bd4), lodestones' 8-step distillation MOSS-Image-SFT-8-steps-dmd (f95df83), and the vae/ of trmz/flux2-klein-alpha (390379b). The LoRA was evaluated with this recipe:

• Trigger: append isolated on a transparent background to the prompt. For a JSON layout prompt, append it to the summary caption.
• Model: the SFT transformer in bf16. The distillation's LoRA (rank 128) and this LoRA are both applied unmerged, computed in fp32 beside each projection and added before the result is rounded back. The distillation's delta is smaller than bf16's rounding step, so merged it would mostly round away.
• Sampler: the SFT model's own step embedder (not the distillation's fp32 replacement), flow-matching timesteps with shift 5, and 8 steps. Each step predicts the clean latent and re-noises it to the next sigma with fresh noise. The distillation's own data-space sampler does not carry the alpha signal: the background comes out as a textured, half-transparent haze.
• Decode: the RGBA VAE, with its own BatchNorm statistics for denormalization, then a 2x2 unpatchify. Above about 1.2 MP, decode in tiles.

pip install "torch>=2.9" "diffusers>=0.38" "transformers>=5.3" safetensors huggingface_hub pillow
python inference_example.py "photo of a red fox sitting, with a big fluffy tail" fox.png --seed 21
python inference_example.py "photo of glowing blue neon tube lettering with the Japanese text 「ラーメン」" neon.png --seed 15 --width 1344 --height 768

The example runs the text encoder first and frees it before the transformer loads, so one 20 GB card is enough. On an RTX A6000 the example reproduces the production runtime's picture byte for byte (the fox above, seed 21). Our production runtime takes 14.9–16.1 s per 1024×1024 picture on an RTX 3080 20GB (text encoder streamed from host memory) and about 10.5 s on an RTX A6000.

Prompts and seeds of the showcase: fennec girl (seed 11, 832×1216), fox (21), sparkling water (3, 832×1216), smoke (4), tea-shop logo (9, a JSON layout with the text 「茶房 みどり」), kanji 「心」 (6), roll cake (12), neon 「ラーメン」 (15, 1344×768).

Training

• Data: 209 RGBA pictures that MOSS-Image itself drew with the recipe above minus this LoRA (bf16 SFT plus the 8-step distillation), on a plain background, then cut out:

  Kind	How the alpha was made
  Subjects, logos, cards and stickers, drawn on flat mid-gray	trmz's Extract-4B LoRA on FLUX.2-klein-base-4B (both Apache-2.0), with a fitted background plate
  Glowing effects and neon, drawn on black	Brightness: alpha is the brightest channel above the black level
  Ink, drawn on white paper	Darkness against the paper, scaled so the darkest ink is opaque

  Every cut-out was checked by eye. The data has no other source: no third-party or scraped images, and no outputs of other image models. Captions are plain English and end with the trigger.

• Recipe: ai-toolkit with trmz's RGBA patch. Rank 32 / alpha 32, AdamW 8-bit, constant learning rate 3e-4, batch 1, 2000 steps, resolutions 512, 768 and 1024, bf16. This release is the step-1750 save; the final save had started to lose its subjects.

Evaluation

12 held-out prompts × 2 seeds through the production runtime, each result checked over a checker and over dark green:

The lettering in the evaluation (「心」, 「茶房 みどり」, 「ケーキセット 680円」) came out right at both seeds.

Limitations

• Shop cards and logo discs come out see-through instead of filled. The training data had few filled panels behind text.
• Smoke and fire have hard edges, and neon lettering has little glow around it.
• Lettering is as good as MOSS-Image's. JSON layout prompts are more reliable than plain prompts.
• Tested only with the recipe above. Other samplers, in particular the distillation's data-space sampler, lose the alpha.

License

Apache-2.0. Everything that went into this LoRA is Apache-2.0 or MIT; see NOTICE. "MOSS-Image" names the model this LoRA is for and does not imply endorsement by its authors.

日本語

MOSS-Image-SFT 用の LoRA です。背景が透明な画像（アルファ付き PNG）を、背景除去の工程なしにサンプラーから直接作れます。被写体の切り抜き、ロゴ、文字、エフェクトに使えます。デコードには trmz の RGBA VAE（trmz/flux2-klein-alpha の vae/）を使います。

• プロンプトの最後に isolated on a transparent background を付けます（JSON レイアウトなら summary の caption に付けます）。
• 手順は上の「How to use it」と inference_example.py のとおりです。20GB の GPU 1 枚で動きます。
• 学習データは MOSS-Image 自身が描いた 209 枚です。切り抜きには Apache-2.0 のモデルとキー処理だけを使い、全部目で確認しました。他社の画像や、他の画像モデルの出力は使っていません。
• 苦手なもの: 塗りつぶしのカードや丸いロゴの円が透けやすいこと、煙や炎の輪郭が硬いこと、ネオンの光の輪が弱いこと。
• ライセンスは Apache-2.0 です（NOTICE を参照）。