IrisKimi-3B-v1

IrisKimi-3B-v1 is an illustration model based on iris-3b-illustration, which is a full fine-tune of Iris-3B. It adds 600 optimizer updates of style tuning, with the adjustment merged into the checkpoint.

This repository provides a complete BF16 image-generation model. The shared Qwen3-VL-4B-Instruct text encoder is obtained separately.

Comparison / 比較

Left to right / 左から: Original Iris 3B, iris-3b-illustration, IrisKimi-3B-v1.

Each row uses the same prompt, negative prompt, seed, and generation settings across all three models. 各行は3モデルで同じプロンプト・ネガティブ・seed・生成設定を使用しています。

Comparison settings / 比較条件

• Resolution: 896 × 1152 per image; comparison sheets show resized previews.
• Steps: 50; CFG: 3.0; flow shift: 4.0; inference precision: BF16.
• Flower garden seed: 20261012.
• City street seed: 20261013.

Flower garden prompt:

an anime girl with green hair and violet eyes wearing a pink dress, standing in a flower garden, upper body, looking at viewer

City street prompt:

an anime man with short silver hair and blue eyes wearing a brown coat, full body, standing on a city street

Negative prompt for both rows:

low quality, blurry, deformed face, bad anatomy, bad hands, extra fingers, text, watermark

Download and use

hf download Aikimi/IrisKimi-3B-v1 IrisKimi-3B-v1.safetensors config.yaml --local-dir models/Iris-3B/checkpoints/IrisKimi-3B-v1

In Aikimi Forge Neo, open the Iris tab, refresh the checkpoint list, and select this model. Prepare the shared Iris runtime and text encoder if needed. The checkpoint uses the official Iris model/config format.

Starting generation settings: approximately 1 megapixel, 50 sampling steps, CFG 3.0, and flow shift 4.0. Use descriptive English prompts and a negative prompt as needed.

日本語

IrisKimi-3B-v1 は、Iris-3Bのフルファインチューニング版である iris-3b-illustration をベースに、600回の更新で絵柄を追加調整し、調整分をマージしたイラスト向けモデルです。

BF16の完全な画像生成モデルとして配布しています。Qwen3-VL-4B-Instructの共通テキストエンコーダーは別途必要です。Aikimi Forge NeoのIrisタブで一覧を更新し、このチェックポイントを選択して使用できます。

License and credits

Apache License 2.0. See LICENSE and NOTICE.

• Original model and architecture: Speridlabs Iris-3B.
• Base illustration model and additional style tuning: Aikimi.
• Separate text encoder: Qwen3-VL-4B-Instruct, under its own license.