Light ConceptsTraining data is a collection of various light concepts I enjoy using that are not overly represented in large datasets, trained as a single lora.<span style="background-color:rgb(26, 27, 30);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">ℹ️ LoRA work best when applied to the base models on which they are trained. Please read the About This Version <span style="background-color:rgb(26, 27, 30);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">on the appropriate base models and workflow/training information.I've trained many of these concepts before, generally they are nice to use to enhance the lighting of generations or give interesting effects.<span style="background-color:rgb(26, 27, 30);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Trained on a large mixed NL and tags dataset, at mixed [1024, 1536] resolutions. Previews are mostly generated at 1024x1536 with a combination of tags and NL prompts.Concept tagsNot limited to, but collected by works containing:<pre><code>
dispersion
hue shifting
refraction
subsurface scattering
translucent

bioluminescence
caustics
dappled moonlight
glowing hot
ultraviolet light</code></pre>Works best in combination with NL if you name a character, describe their basic appearance, and finish with descriptions of light sources and their effects on the scene:<pre><code>
A vibrant and dynamic illustration of Hoshimachi Suisei from Hololive, featuring her squatting in front of a glowing triangular prism.
A beam of white light enters the prism from the left and refracts into a vibrant rainbow on the right. The background is a solid dark grey to emphasize the lighting effects.</code></pre>

Light Concepts

Trained on <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima">Anima Preview</a> (version 1)Assume that any lora trained on the preview version won't work well on the final version. Trained the mixed tags and natural language dataset.Config:<pre><code># dataset-anima.toml
# Resolution settings.
resolutions = [1024, 1280]

# Aspect ratio bucketing settings
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7

[[directory]] # IMAGES
# Path to the directory containing images and their corresponding caption files.
path = '/mnt/d/training_data/images'
num_repeats = 1
resolutions = [1024, 1280]

# Change these paths
output_dir = '/mnt/d/anima/training_output'
dataset = 'dataset-anima.toml'

# training settings
epochs = 50
micro_batch_size_per_gpu = 16
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 100
train_llm_adapter = true

# eval settings
eval_every_n_epochs = 1
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1

# misc settings
save_every_n_epochs = 1
checkpoint_every_n_minutes = 120
activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 1
steps_per_print = 1

[model]
type = 'anima'

transformer_path = '/mnt/c/workspace/models/diffusion_models/anima-preview.safetensors'
vae_path = '/mnt/c/workspace/models/vae/qwen_image_vae.safetensors'
qwen_path = '../qwen0.6/Qwen3-0.6B/'
dtype = 'bfloat16'
timestep_sample_method = 'logit_normal'
sigmoid_scale = 1.0
shift = 3.0

# Caption Processing Options
cache_text_embeddings = false
# NOTE: Requires cache_text_embeddings = false to work!
# For cached embeddings, use cache_shuffle_num in your dataset config instead.
shuffle_tags = true
tag_delimiter = ', '
keep_first_n_tags = 5
shuffle_keep_first_n = 5
tag_dropout_percent = 0.3
protected_tags_file = './protected_tags.txt'

nl_shuffle_sentences = false
nl_keep_first_sentence = true

# 'tags' 'nl' 'mixed'
caption_mode = 'mixed'

debug_caption_processing = false
debug_caption_interval = 1000

[adapter]
type = 'lora'
rank = 64
dtype = 'bfloat16'

# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
[optimizer]
type = 'adamw_optimi'
lr = 8e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8</code></pre>

Files with SHA256: f0a3feb853cb77391fefbbf3a892c54b21963ef57f6bd80b5161e3f8462fedfa