1. Overview
Flux-1-dev is a two-dimensional animation style LoRA (Low-Rank Adaptation) trained by base model. It focuses on retaining the original character structure of the base model while enhancing the vividness and three-dimensional sense of lines and colors, taking into account the restoration of hand details and the generalization ability of multiple scenes. It is suitable for various animation generation needs such as character standing drawings, background interactions and dynamic actions.
2. Core features
Controllable style concentration 0.6–0.7: Mild filter, retaining the details of the original model 0.7–0.8: Moderate style, the best balance between consistency and details 0.8–0.9: Rich color matching and light and shadow, stronger illustration texture 0.9–1.0: Extreme filter, suitable for experiments and special effects
Hand stability: Output clear and complete five fingers and natural gestures in common actions (reading, stage performances, fighting postures, etc.)
Multi-scene adaptation: It has been verified that smooth style transfer can be maintained in 8 major themes such as campus, night battle, magic, mecha, and horror
Lightweight and efficient: LoRA file ≈ 15 MB, only occupies ∼5% of the base model video memory during inference, and can be stacked with other LoRA (such as RefinedHands)
3. Usage
Recommended parameters Resolution: 1024×1024 Sampling: DPM++ 2M Karras (28 steps, CFG 7.5) clip_skip: 2
Negative prompts: (worst quality, lowres:1.2), watermark, signature, badhand, extra fingers
Weight adjustment Daily creation: 0.7–0.8 Illustration texture: 0.8–0.9 Light filter: 0.6–0.7 Special effects/experiments: ≥ 0.9
IV. Example Prompt
1girl, dynamic pose, cherry blossom petals, detailed lineart, soft bloom bokeh, e08_style @0.8
V. Notes
In high-weight scenes, random accessories (cat ears, wreaths) may appear, which can be suppressed with the negative prompt no animal ears
If you want to fine-tune the light and shadow or color level, you can add modifiers such as cinematic lighting and vibrant colors in the prompt