# Skill: Produce top-quality, style-consistent images with FLUX.2

> Generate clean, "expensive"-looking images and keep a consistent style/character across a series using FLUX.2 (Black Forest Labs); pick the dev or klein variant by quality vs hardware, and respect its commercial license.

Source: sodigi-learn - comfyui/flux-2 - https://sodigi.io/learn/comfyui/flux-2

## When to use

- You need a polished, professional image with strong prompt adherence and few artifacts.
- You need a consistent look/character across multiple images (brand, cover series, campaign).
- You want a fast model to iterate many variants, or your GPU is mid-range (use klein).
- The image has little or no text (for text-heavy images use Qwen-Image instead).

## Core rules

- FLUX.2 dev (~32B) = maximum quality + multi-reference consistency (feed several references to hold one style/character across a series).
- FLUX.2 klein (4-9B) = fast, ~13GB VRAM on the smallest, unified generate+edit; for iteration and weaker hardware.
- Use FLUX.2 when quality matters and text on the image is minimal; route text-in-image work to Qwen-Image.
- License is NOT Apache: commercial use of the larger models generally needs a separate license - check it BEFORE selling outputs or shipping in a paid product.
- For a series in one style, give dev reference images rather than re-describing the style each time.
- If dev does not fit in VRAM, use klein or a GGUF/quantized build.
- For even larger scenes / 1000-word prompts, consider HunyuanImage 3.0 (80B MoE, heavy GPU).

## Procedure

1. Choose variant: dev for final hero images and consistency; klein for speed/iteration or mid GPU.
2. Download the chosen FLUX.2 model (or a GGUF build if VRAM is tight) and load it.
3. For a consistent series, attach reference images so dev preserves style/character.
4. Write a precise prompt describing the scene (FLUX follows prompts closely).
5. Before any commercial use, read and comply with the FLUX license.

## Ready-to-use prompt

```
Generate a high-quality image with FLUX.2 (use dev for the final, klein for quick drafts).
Scene: <detailed description>
Style/consistency: <if a series, "match the attached reference images for style and character">
Constraints: photorealistic/illustrated as specified, clean hands/faces, follow the prompt exactly.
Note: keep on-image text minimal (route heavy text to Qwen-Image). Confirm licensing if this output is commercial.
```

## Pitfalls

- Using FLUX.2 for text-in-image work (use Qwen-Image).
- Running dev on a weak GPU instead of klein or a GGUF build (out-of-memory).
- Ignoring the FLUX license before selling outputs - it is not Apache 2.0.
- Re-describing the style for each image instead of giving dev reference images for consistency.
- Reaching for HunyuanImage's 80B when FLUX.2 already covers the task at far lower VRAM.