stixxert/topdown-medieval-pixelart

Dataset Card for Top-Down Medieval Pixel Art

Quick Facts

Dataset Description

Top-Down Medieval Pixel Art is a synthetic dataset of 100 top-down (overhead) medieval fantasy pixel-art game assets, each paired with a descriptive natural-language caption. It was designed for LoRA fine-tuning of diffusion models to learn a consistent top-down camera angle, decoupled from visual style as much as possible.

Each image is a 1024×1024 PNG with a white background, rendered in a crisp 16×16 pixel-art aesthetic. Asset categories include medieval structures (buildings, towers, workshops, outposts), vegetation (brambles, reed clusters), and terrain (boulders, rock formations).

You may notice that the dataset is dominated by structures. This was not by design — the generation pipeline simply produced better results for buildings than for vegetation or terrain (see Curation Rationale). Despite the skew, the LoRA trained on this dataset successfully transfers top-down perspective to categories it barely saw during training, including character sprites and terrain features.

• —Curated by: stixxert
• —Language(s): English (captions)
• —License: Non-commercial use only — see License for full dependency chain

Dataset Sources

• —Base model: black-forest-labs/FLUX.2-dev
• —Camera-angle LoRA: lovis93/Flux-2-Multi-Angles-LoRA-v2
• —Pixel-art post-processing: dimtoneff/ComfyUI-PixelArt-Detector (MIT)
• —Background removal: Loewen-Hob/rembg-comfyui-node-better (MIT, based on rembg)

Uses

Direct Use

This dataset is intended for:

• —LoRA adapter training — fine-tuning diffusion models to produce consistent top-down camera angles on generated images
• —Game-development asset pipelines — generating consistent sprites, tilesets, and props with overhead perspective
• —Camera-pose conditioning research — studying how diffusion models learn and apply viewpoint concepts from training data
• —Style-transfer experiments — exploring pixel-art aesthetics in generative image models

The dataset is specifically designed to pair with FLUX.2 Klein 4B using the Ostris AI Toolkit.

This dataset was used to train the StrategAI LoRA adapters — six LoRA variants for FLUX.2 Klein 4B that enforce top-down medieval style in generated game assets. See the companion Model Card for training details and evaluation results.

Out-of-Scope Use

This dataset is not suitable for:

• —Commercial asset production — inherits non-commercial restriction from FLUX.2 [dev]; see License for the full chain
• —Photorealistic generation — all images are pixel art; the dataset will not teach a model to produce photorealistic top-down views
• —Training models intended to violate upstream licenses — any use must comply with FLUX.2 [dev] (non-commercial) and Flux-2-Multi-Angles-LoRA-v2 (creativeml-openrail-m) terms
• —Generating NSFW, harmful, or deceptive content
• —Identity or biometric inference — the dataset contains no human subjects or personally identifiable information

License

This dataset is a derivative work built from multiple upstream components. Every image and caption inherits license obligations from the models and tools used to create it.

Upstream License Chain

What This Means for You

• —Non-commercial use only. The most restrictive upstream license (FLUX.2 [dev]) prohibits commercial use. This restriction flows through to all images and captions in this dataset.
• —You must retain the FLUX.2-dev license when redistributing this dataset or models trained on it.
• —You must comply with CreativeML Open RAIL-M usage restrictions (no harmful or illegal use, no medical advice, no law enforcement use, etc.) when using the Multi-Angles LoRA trained weights.
• —The MIT-licensed tools (PixelArt-Detector, rembg) impose no additional restrictions on the output images.
• —If you hold a separate commercial license for FLUX.2 [dev] from Black Forest Labs, the non-commercial restriction on this dataset may no longer apply — consult your licensing agreement.

Summary: The effective license of this dataset is non-commercial, with Open RAIL-M usage restrictions. You may not use this dataset for commercial purposes unless you hold a separate commercial license for FLUX.2 [dev].

Dataset Structure

File Layout

images/
├── 00001.png
├── 00002.png
├── ...
├── 00100.png
└── metadata.jsonl

Each .png file is a 1024×1024 pixel RGBA image on a white background. The companion metadata.jsonl contains one JSON object per line with the following schema:

Data Splits

This dataset has no train/test/validation split. All 100 images form a single set intended for LoRA training. Users should create their own splits if needed for evaluation.

Example Data Point

{
  "id": "00009",
  "file_name": "00009.png",
  "text": "top-down view. A medieval granary in a steampunk industrial expansion setting, featuring brass fittings and riveted mechanical accents, ornate but functional character, ornate detail, simple footprint, aged limestone blocks, weathered, smoky sepia palette",
  "asset_family": "structure"
}

Dataset Statistics

Dataset Creation

Curation Rationale

This dataset was created to test a specific hypothesis in LoRA fine-tuning:

A low-rank adapter can learn a compositional concept like camera angle from ~100 examples if the training data is consistent in perspective and diverse in subject matter.

The target use case is game-development pipelines where dozens or hundreds of assets must share a consistent top-down perspective. Rather than training a LoRA on each asset type separately, the goal is a single adapter that enforces the overhead camera angle regardless of what is being depicted.

Key design decisions:

The heavy skew toward structure (97 of 100) was not intentional — it reflects where the generation pipeline worked well and where it struggled. The teacher pipeline (FLUX.2 [dev] + Multi-Angles LoRA v2, pixel-art quantization, background removal) reliably produced clean, readable structures. But the same pipeline faltered on non-structure categories: vegetation blurred into indistinct masses after quantization; rock formations rarely survived the strict top-down constraint with convincing geometry. Each usable vegetation or terrain image required many more generation attempts than a structure image. Within a curation budget of roughly 150–200 raw generations, the practical outcome was a structure-dominated dataset.

The skew has a silver lining: it provides a stronger test of concept isolation. If a LoRA trained on 97% structures successfully transfers top-down perspective to unseen categories (0% character sprites, 3% terrain), the angle concept has been genuinely isolated from subject matter.

Source Data

Data Collection and Processing

All images in this dataset are synthetically generated. No photographs, hand-drawn art, or web-scraped images were used.

The generation pipeline uses ComfyUI with the following components:

The camera angle was enforced at generation time by the Multi-Angles LoRA, not by prompt engineering. This ensures every image truly is top-down — there is no variation in perspective.

Processing steps:

1. 1.Prompt generation — diverse medieval fantasy captions are produced from structured templates
2. 2.Image generation — ComfyUI runs FLUX.2 [dev] + Multi-Angles LoRA with each prompt via HTTP API
3. 3.Pixel-art quantization — ComfyUI-PixelArt-Detector reduces the color palette to a 16×16 pixel-art aesthetic
4. 4.Background removal — rembg strips the generated background, replacing it with solid white
5. 5.Manual curation — a human reviewer deletes images with artifacts, perspective errors, or poor composition; ~150–200 raw images were generated and culled to the best 100
6. 6.Dataset packaging — embedded metadata is stripped, files are renumbered, and the final metadata.jsonl is produced

Who are the source data producers?

The images were generated by the FLUX.2 [dev] diffusion model (Black Forest Labs), conditioned by prompt templates authored by the dataset curator. The camera angle was enforced by the Flux-2-Multi-Angles-LoRA-v2 adapter (lovis93).

No human artists, photographers, or annotators were involved in the image creation process. No web-scraped or crowd-sourced data was used.

Annotations

Annotation Process

Captions were machine-generated via structured prompt templates that combine:

• —Asset type (granary, watchtower, sawmill, barracks, etc.)
• —Setting/context (steampunk industrial, feudal military, mystical arcane, etc.)
• —Architectural features (brass fittings, rune-carved stone, timber bracing, etc.)
• —Condition (weathered, war-damaged, well-maintained, etc.)
• —Color palette (ash-and-ember, mystic violet-cyan, cool medieval, etc.)

Each caption is prefixed with the angle phrase top-down view. to bind the overhead perspective in the text embedding space.

Captions are natural-language sentences — not structured tags, booru-style keywords, or comma-separated token lists. This format was chosen for compatibility with FLUX.2's T5 text encoder, which expects natural language input.

Three caption variants were derived from each original caption for training experiments:

These caption variants are used to test the hypothesis that less text detail leads to purer angle-concept learning. They are not included in this published dataset.

Who are the annotators?

Captions were generated using structured templates. The template design and caption review were performed by the dataset curator. No crowd-sourced or third-party annotators were involved.

Personal and Sensitive Information

This dataset contains no personal or sensitive information:

• —No human subjects, faces, or biometric data
• —No names, addresses, or personally identifiable information
• —No geolocation data
• —No financial, health, or demographic data
• —All images are synthetic renderings of imaginary medieval structures and landscapes

Bias, Risks, and Limitations

Known Biases

• —Medieval/fantasy domain bias — all images depict medieval fantasy assets. A LoRA trained on this dataset will primarily produce medieval-themed outputs when the trigger token is activated.
• —Pixel-art style bias — all images use a 16×16 pixel-art aesthetic. The LoRA may inadvertently learn pixel-art style alongside camera angle, despite efforts to isolate the angle concept.
• —Structure-heavy distribution — 97% of images are structures. This reflects generation pipeline constraints (see Curation Rationale) rather than a design choice. The camera-angle concept may not transfer as reliably to vegetation, terrain, or organic subjects.
• —Single perspective — all images are strictly top-down. The LoRA will not learn isometric, side-view, or ¾-angle perspectives.

Technical Limitations

• —Small dataset size — 100 images is sufficient for low-rank LoRA training but may not support full model fine-tuning or high-rank adapters without overfitting
• —Synthetic origin — images were generated by a diffusion model and may contain artifacts typical of AI-generated imagery
• —Checkpoint dependence — the dataset was generated with specific versions of FLUX.2 [dev] and ComfyUI custom nodes; reproducing the exact images requires the same software versions
• —No negative examples — the dataset contains only top-down images with no counter-examples of other angles, which may limit the LoRA's ability to suppress non-top-down perspectives

Recommendations

• —Use a trigger token — set trigger_word in your training config (e.g., <tdp>). The Ostris AI Toolkit prepends it automatically. Always include both the trigger token and the angle phrase at inference: <tdp> top-down view. a medieval watchtower...
• —Pair with FLUX.2 Klein 4B — the dataset was designed for and tested with this specific base model
• —Evaluate checkpoints — train with frequent checkpoint saves (every 200 steps) and evaluate against out-of-distribution prompts to detect overfitting to individual assets
• —Consider caption detail level — experiment with detailed, minimal, and ultra-minimal captions to find the right balance between style isolation and generation quality
• —Do not use commercially — respect the upstream license chain: FLUX.2 [dev] (non-commercial) and Flux-2-Multi-Angles-LoRA-v2 (creativeml-openrail-m). See License for details.

Citation

If you use this dataset in published research, please cite both the dataset and the underlying models.

BibTeX:

@misc{topdown-medieval-pixelart-2026,
  author = {stixxert},
  title = {Top-Down Medieval Pixel Art},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/stixxert/topdown-medieval-pixelart}},
  note = {Synthetic dataset of 100 top-down medieval pixel-art game assets for LoRA fine-tuning}
}

@misc{flux2-dev,
  author = {Black Forest Labs},
  title = {FLUX.2 [dev]},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/black-forest-labs/FLUX.2-dev}}
}

@misc{multi-angles-lora,
  author = {lovis93},
  title = {Flux-2-Multi-Angles-LoRA-v2},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/lovis93/Flux-2-Multi-Angles-LoRA-v2}}
}

APA:

stixxert. (2026). Top-Down Medieval Pixel Art [Dataset]. Hugging Face. https://huggingface.co/datasets/stixxert/topdown-medieval-pixelart

Glossary

• —Trigger token — a unique token (e.g., <tdp>) prepended to captions at training time to activate the LoRA at inference. Not present in the published dataset; users choose their own.
• —Angle phrase — the text prefix top-down view. that binds the overhead perspective in the text embedding space. Present in all published captions.
• —LoRA (Low-Rank Adaptation) — a parameter-efficient fine-tuning method that trains small adapter matrices rather than updating all model weights.
• —Camera angle / perspective — the viewpoint from which a scene is rendered. "Top-down" means the camera is positioned directly above the subject, looking straight down.
• —Ostris AI Toolkit — an open-source training toolkit for FLUX models, used to train the LoRA adapters in the companion project.
• —Pixel-art quantization — reducing an image's color palette to produce a retro pixel-art aesthetic (16×16 pixel grid in this dataset).

Dataset Card Authors

stixxert