MODEL_CARD.md

# Model Card for StyleController

## Model summary

StyleController is a controllable image-stylization method built on image-editing diffusion backbones. It combines a full-strength LoRA endpoint with lightweight projector anchors learned at discrete style strengths. Interpolation between these anchors produces a continuous stylization path while aiming to preserve the content image's structure.

This Hugging Face repository is organized as a multi-variant weight collection. Each top-level model directory is self-contained with respect to adapter weights, projector weights, machine-readable configuration, and checksums. Base-model weights and executable source code are not duplicated here.

## Current release

### qwenimage_edit_2509

| Property | Value |
| --- | --- |
| Runtime base model | `Qwen/Qwen-Image-Edit-2509` |
| Relationship | Custom LoRA adapter with strength projectors |
| Tensor format | Safetensors |
| Tensor dtype | BF16 |
| LoRA rank | 128 |
| Transformer layers | 0–59, inclusive |
| Target-module groups | 10 per transformer layer |
| Projector type | `low_rank_linear` |
| Stored anchors | 0.1 through 0.9 |
| Virtual endpoints | 0.0 zero projector; 1.0 identity projector |
| Main weight parameters | 613,416,960 |
| Projector parameters | 9,907,200 per anchor |
| Total stored parameters | 702,581,760 |
| Total weight size | 1,406,700,168 bytes |

## Intended use

The release is intended for:

- research on reference-based image stylization;
- controllable image-to-image generation;
- continuous style-strength interpolation;
- evaluation and reproduction of the StyleController method;
- non-deceptive creative workflows where users have the rights needed for the input images and requested outputs.

The adapters are not intended to identify individuals, authenticate media, make high-stakes decisions, or provide guarantees of copyright clearance for generated images.

## How to use

Use the inference entry point in the [StyleController source repository](https://github.com/ReyChiaro/StyleController). A complete command is provided in the repository-level [`README.md`](README.md). The essential compatibility requirements are recorded in [`qwenimage_edit_2509/stylecontroller_config.json`](qwenimage_edit_2509/stylecontroller_config.json).

The scalar strength is expected to lie in `[0, 1]`. Projector checkpoints `s01` through `s09` correspond to strengths `0.1` through `0.9`. The source implementation constructs both missing endpoints and interpolates projector parameters at the requested strength.

## Training data

StyleController was trained with the SmoothStyle dataset. SmoothStyle contains content images, style references, generated stylization targets, and scalar strength supervision.

The dataset has mixed provenance and its licensing is separate from this model-weight license. Users should consult the current [`ReyChiaro/SmoothStyle`](https://huggingface.co/datasets/ReyChiaro/SmoothStyle) dataset card, per-asset provenance records, third-party notices, and release status before using or redistributing dataset assets. The Apache-2.0 license for these adapter files does not grant rights to third-party training or input images.

## Evaluation

The project evaluates stylization fidelity, content preservation, and smoothness across strength trajectories. Final, release-specific quantitative tables and paper citation metadata have not yet been included in this weight repository. Do not treat the absence of a reported metric as evidence of performance on untested domains.

## Safety and responsible use

Users are responsible for obtaining appropriate rights and consent for content and style references. Avoid deceptive impersonation, non-consensual identity manipulation, removal or falsification of provenance information, and attempts to imitate living artists in jurisdictions or contexts where doing so would violate applicable rights or platform rules.

Generated outputs should be reviewed before publication. For sensitive workflows, preserve input provenance, prompts, model/config revisions, random seeds, and output metadata.

## Reproducibility and integrity

- Configuration: `qwenimage_edit_2509/stylecontroller_config.json`
- Repository index: `model_manifest.json`
- Weight checksums: `qwenimage_edit_2509/checksums.sha256`
- Source code: <https://github.com/ReyChiaro/StyleController>
- Base model: <https://huggingface.co/Qwen/Qwen-Image-Edit-2509>

The Safetensors files contain tensor data only and currently carry no embedded metadata. External configuration files are therefore part of the release contract and should be distributed with the weights.

## License and attribution

The adapter and projector weight files are released under Apache-2.0. StyleController source code is licensed separately under MIT. Qwen-Image-Edit-2509 is distributed by its authors under Apache-2.0 and is not included in this repository. See [`LICENSE`](LICENSE) and [`NOTICE`](NOTICE).

## Contact

Please use the [StyleController GitHub issue tracker](https://github.com/ReyChiaro/StyleController/issues) for technical problems, model-card corrections, provenance concerns, and takedown requests until a dedicated project contact is published.