README.md

---
license: apache-2.0
base_model:
- black-forest-labs/FLUX.2-klein-4B
pipeline_tag: text-to-image
tags:
- flux,
- flux2,
- lora,
- text-to-image,
- image-generation,
- style,
- black-and-white,
- chiaroscuro,
- ai-toolkit
---

# GustaveDoreStyle LoRA

GustaveDoreStyle is a FLUX.2 LoRA trained to reproduce the visual language of 19th-century black-and-white engraving: dense cross-hatching, dramatic chiaroscuro, monumental composition, deep shadows, and strong narrative scale.

The model is intended for modern reinterpretation: science fiction, fantasy, architecture, mythic scenes, cinematic environments, and contemporary subjects rendered with a Gustave Dore-inspired engraving aesthetic.

## Trigger Word

Use this trigger near the beginning of your prompt:

```txt
gustave_dore_style
```

Recommended prompt structure:

```txt
gustave_dore_style, [subject], black and white engraving, dense cross hatching, dramatic chiaroscuro, monumental scale, epic composition
```

## Example Prompts

```txt
gustave_dore_style, a lone astronaut standing before a cathedral-sized alien machine, black and white engraving, dense cross hatching, dramatic chiaroscuro, monumental scale, epic composition
```

![A lone astronaut before a cathedral-sized alien machine](examples/example-01-astronaut-alien-machine.jpg)

```txt
gustave_dore_style, a cybernetic prophet walking through a ruined city, black and white wood engraving, intricate linework, deep shadows, divine light breaking through clouds
```

![A cybernetic prophet walking through a ruined city](examples/example-02-cybernetic-prophet.jpg)

```txt
gustave_dore_style, a tiny human figure facing a colossal ancient robot, black and white engraving, extreme scale contrast, dramatic light, dense linework
```

![A tiny human figure facing a colossal ancient robot](examples/example-03-ancient-robot.jpg)

```txt
gustave_dore_style, a colossal ancient tree towering over a tiny group of explorers, black and white wood engraving, dense cross hatching, dramatic chiaroscuro, extreme scale contrast, intricate linework, dark atmospheric forest, beams of light breaking through branches
```

![A colossal ancient tree towering over tiny explorers](examples/example-04-colossal-tree.jpg)

```txt
gustave_dore_style, a lone figure standing on a cliff facing a vast storm over an endless ocean, black and white engraving, dramatic sky, heavy clouds, powerful wind, deep shadows, monumental composition, detailed waves, epic scale
```

![A lone figure on a cliff facing a vast storm](examples/example-05-storm-cliff.jpg)

```txt
gustave_dore_style, a massive underground cavern filled with towering pillars and a hidden glowing structure in the distance, black and white engraving, dense cross hatching, dramatic lighting from below, intricate textures, tiny human silhouettes for scale, mysterious atmosphere
```

![A massive underground cavern with towering pillars](examples/example-06-underground-cavern.jpg)

## Recommended Settings

| Parameter | Recommendation |
|---|---:|
| Base model | `black-forest-labs/FLUX.2-klein-base-4B` |
| Resolution | `1024x1024` or similar |
| Steps | `20-30` |
| Guidance / CFG | `4-7` |
| Sampler | Flowmatch-compatible sampler |
| LoRA strength | `0.7-1.0` |

Start around `0.8` strength. Raise it if you want a stronger engraving look; lower it if the style overwhelms the subject.

## Prompting Tips

This LoRA responds best to clear composition language. Useful concepts:

- tiny figure versus enormous structure
- cathedral scale
- deep foreground and distant background
- dramatic sky
- divine light or beams of light
- black-and-white engraving
- wood engraving
- dense cross hatching
- intricate linework
- deep shadows
- monumental composition

Avoid overly vague prompts. The style is strongest when the scene has a clear subject, lighting direction, and sense of scale.

## Training Details

| Attribute | Value |
|---|---|
| Training type | LoRA |
| Base model | `black-forest-labs/FLUX.2-klein-base-4B` |
| Architecture | `flux2_klein_4b` |
| Training tool | Ostris AI Toolkit |
| Dataset size | 237 curated image/caption pairs |
| Dataset resolution bucket targets | 512, 768, 1024, 1280 |
| Trigger token in captions | `gustave_dore_style` |
| Network rank | 32 |
| Network alpha | 32 |
| Conv rank | 16 |
| Conv alpha | 16 |
| Optimizer | `adamw8bit` |
| Learning rate | `0.0001` |
| Batch size | 1 |
| Gradient accumulation | 4 |
| Noise scheduler | `flowmatch` |
| Training dtype | `bf16` |
| Text encoder training | Disabled |

The training run saved checkpoints every 250 steps. The visible style begins to converge early, with stronger results around the later checkpoints.

## Dataset Curation

The dataset was curated from Gustave Dore-style biblical engraving material and cleaned for style training.

Removed or reduced:

- text-heavy scans
- page borders
- page artifacts
- low-quality images
- weak compositions

Focused on:

- full compositions
- dramatic lighting
- complex environments
- figure and scale relationships
- dense linework
- high-contrast black-and-white engraving structure

## Use Cases

- Science-fiction scenes in classical engraving style
- Fantasy illustration
- Biblical or mythological reinterpretation
- Concept art exploration
- Narrative/cinematic stills
- Stylized animation and previsualization
- Architectural and monumental environment studies

## Notes

This LoRA is not only a texture filter. It tends to influence composition, lighting, and scale. The best results come from prompts that describe spatial hierarchy and dramatic contrast, not only surface style.

If the model output is too literal or too dark, reduce LoRA strength slightly and add clearer subject and lighting language.

## Example Image Files

The example images used above are expected in this repository path:

```txt
examples/
  example-01-astronaut-alien-machine.jpg
  example-02-cybernetic-prophet.jpg
  example-03-ancient-robot.jpg
  example-04-colossal-tree.jpg
  example-05-storm-cliff.jpg
  example-06-underground-cavern.jpg
```

Upload the `examples` folder together with this README so the prompt/image pairs render correctly on Hugging Face.

## Repository

Model page:

```txt
https://huggingface.co/Neofuturism/GustaveDoreStyle
```