Illuminate — SDXL Illumination Slider

Illuminate is an experimental bidirectional concept-slider LoRA for SDXL designed to alter the apparent illumination of an image.

Positive and negative LoRA weights move generation in opposing illumination directions while generally retaining the underlying subject and composition. It is intended as a practical generation control as well as an experiment in LoRA-based semantic steering.

Developed by DetroitWobbly.

What It Does

Illuminate was trained around a deliberately simple semantic contrast:

• Positive pole: a woman in brightness
• Negative pole: a woman in darkness
• Target class: a woman

The resulting LoRA can affect illumination beyond the training subject class and has been tested on people, environments, and img2img inputs.

No trigger word is required.

Weight Range

A useful working range is approximately:

-4 to +4

Start near ±1 and increase magnitude progressively.

Higher absolute weights produce stronger effects but can also introduce collateral changes to contrast, color, rendering, or scene details. Exact behavior varies with checkpoint, prompt, seed, and other loaded LoRAs.

Model Details

• Model type: SDXL concept-slider LoRA
• Developer: DetroitWobbly
• Architecture: Stable Diffusion XL LoRA
• Training framework: AI Toolkit
• Training process: concept_slider
• Training base: Juggernaut XL Ragnarok
• Base checkpoint: juggernautXL_ragnarokBy.safetensors
• Base SHA-256: DD08FA32F98D05A2443CA1419E46DF1575A0811F6E3B246D9DD47FF20F5EB66A
• Trigger word: None
• Primary function: Bidirectional illumination steering

Usage

Load Illuminate as an ordinary SDXL LoRA.

Example:

<lora:Illuminate_v1:1.0>

Try the same prompt and seed at several weights:

-4, -2, -1, 0, +1, +2, +4

The zero-weight generation provides a useful baseline for evaluating the effect.

Illuminate also works in img2img workflows. Lower denoise strengths can be useful when the goal is to alter illumination while retaining an existing composition.

Training Details

Illuminate was trained using AI Toolkit's concept-slider training process.

Slider Definition

• Positive prompt: a woman in brightness
• Negative prompt: a woman in darkness
• Target class: a woman
• Guidance strength: 3
• Anchor strength: 1
• Anchor class: None

LoRA Network

• Linear rank: 32
• Linear alpha: 32
• Convolutional rank: 16
• Convolutional alpha: 16

Training

• Steps: 3,000
• Batch size: 1
• Gradient accumulation: 1
• Optimizer: AdamW8bit
• Learning rate: 1e-4
• Weight decay: 1e-4
• Training precision: BF16
• Loss: MSE
• Noise scheduler: DDPM
• Timestep sampling: sigmoid
• Content/style setting: balanced
• Gradient checkpointing: enabled
• U-Net training: enabled
• Text encoder training: disabled
• EMA: disabled
• Training resolutions: 512, 768, 1024
• Caption dropout: 0.05

Intermediate checkpoints and samples were saved every 250 steps.

Training-Time Probe

Training samples used the fixed seed 42 with the simple prompt a woman at LoRA multipliers:

• -1.0
• 0.0
• +1.0

Samples used 30 steps and guidance scale 6.

This provided a fixed reference for observing development of the slider during training without adding illumination language to the inference prompt.

Evaluation

Illuminate has been tested using controlled weight sweeps in which prompt, seed, and generation settings are held constant while LoRA strength is varied.

Testing has included:

• human subjects
• environmental scenes
• out-of-domain prompts
• img2img inputs
• positive and negative weight sweeps
• weights substantially beyond ±1
• interaction between illumination changes and preservation of the original subject/composition

Observed behavior suggests a practical range of approximately -4 to +4, although the useful range depends on the base checkpoint and generation context.

Evaluation focuses not only on whether illumination changes, but also on collateral effects including:

• subject preservation
• composition preservation
• contrast
• shadow/highlight behavior
• color and tonal shifts
• morphology
• background changes
• artifacts at extreme weights

Further systematic evaluation is being developed using the Checkpoint Evaluation Framework (CEF).

Limitations

Illuminate is a learned LoRA control rather than a deterministic lighting operator.

Its effect depends on the capabilities and priors of the SDXL checkpoint with which it is used. Results can vary substantially with prompt, checkpoint, CFG, sampler, other LoRAs, and inference weight.

At stronger weights, illumination changes may be accompanied by changes in color, contrast, rendering, background details, or other image characteristics.

The slider should therefore not be interpreted as a perfectly isolated mathematical lighting axis.

Cross-checkpoint compatibility is experimental and should be evaluated independently.

Research Notes

Illuminate is part of a series of experiments investigating whether broad visual concepts can be represented as useful bidirectional controls using LoRA adapters.

Questions of interest include:

• How cleanly does the learned concept separate from unrelated image properties?
• How monotonic is the effect across inference weights?
• How far beyond conventional LoRA weights does useful steering persist?
• How well does the control transfer to SDXL checkpoints other than its training base?
• Does the learned direction generalize beyond the subject class used to define the training poles?
• How does the slider interact with style and character LoRAs?

Observed behavior is reported behaviorally and should not be interpreted as proof that the adapter contains a perfectly isolated internal semantic direction.

Author

DetroitWobbly

Independent experiments in SDXL, LoRA training, concept sliders, semantic steering, automated model evaluation, and vision-model-assisted analysis.