TUaxhuiax/nimbus-theme-generator

Nimbus Theme Generator

nimbus-theme-generator is a LoRA adapter for Stable Diffusion XL that turns short design briefs into complete design-system themes: color tokens, spacing scales, typography ramps and component states.

Highlights

• Generates a full tokens.json bundle from a one-line brief such as "calm fintech dashboard, dense spacing, high contrast".
• Keeps semantic token names stable across runs (e.g. color.surface.raised).
• Ships with a token_linter that rejects inaccessible contrast pairs.
• Trained on the companion Nimbus UI Screens dataset.

Quick start

from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
pipe.load_lora_weights("TUaxhuiax/nimbus-theme-generator", weight_name="nimbus_theme_lora.safetensors")

Intended use

Use case	Supported	Notes
Design-system bootstrap	yes	Best results with 3-8 keywords
Brand re-skinning	yes	Provide an existing tokens.json as seed
Photorealistic assets	no	Not a general image model

Training details

• Base model: stabilityai/stable-diffusion-xl-base-1.0
• Adapter rank: 16, alpha 32
• Steps: 12,000 @ batch size 4
• Precision: bf16 with gradient checkpointing

Evaluation (CLIP score on 2,000 held-out briefs)

Split	CLIP score	Token validity
in-domain	0.312	99.4%
out-of-domain	0.274	96.1%

Limitations

• Occasional snake_case token drift for compound words. Compound names are split on camel-case boundaries before rendering.
• Large palettes (> 24 colors) may be truncated.
• Does not emit motion or interaction tokens yet.

Citation

@misc{nimbus_theme_generator,
  title  = {Nimbus Theme Generator},
  author = {Nimbus Design Kit},
  year   = {2026}
}

Contact

Questions? Open an issue in the design-tools tracker and include the brief you used (keep underscores intact).