Jhenderson112/lorabench-freya-lindgren

Freya Lindgren — SDXL LoRA

A single-persona photoreal LoRA for Stable Diffusion XL trained on RealVisXL_V4.0. Generates consistent headshots and portraits of a recurring character — a 32-year-old Swedish woman with platinum-blonde wavy hair, light blue eyes, and natural freckles.

This is the 8th of 8 women in the G-3 persona pipeline; one of the flagship characters in the Lorabench consumer-validation queue.

Example gallery

Before / without LoRA (same prompt, same seed, trigger removed):

After / with LoRA (trigger Freya_Lindgren first):

Trigger

Use the token Freya_Lindgren (with the underscore) at the start of your prompt. The model was trained with shuffle_caption=true and keep_tokens=1, so the trigger can be in any position but should appear first for best results.

Recommended settings

Setting	Value
LoRA strength (UNet + CLIP)	0.85
Sampler	Euler
Scheduler	normal
Steps	22
CFG	7.0
Resolution	1024 × 1024
Negative prompt	lowres, bad anatomy, bad hands, watermark, text, jpeg artifacts, blurry, overexposed, plastic skin, ugly, deformed, extra limbs

Training details

Field	Value	Source
Base model	RealVisXL_V4.0.safetensors	TOML [model]
Dataset	25 captioned images	TOML [dataset]
Resolution	1024×1024	TOML [training]
Network module	networks.lora	TOML [network]
Network dim (rank)	32	TOML [network]
Network alpha	16	TOML [network]
Optimizer	AdamW8bit	TOML [optimizer]
Learning rate	1e-4	TOML [optimizer]
UNet LR	1e-4	TOML [optimizer]
Text encoder LR	5e-5	TOML [optimizer]
LR scheduler	cosine_with_restarts	TOML [optimizer]
LR scheduler cycles	3	TOML [optimizer]
LR warmup steps	100	TOML [optimizer]
Max train epochs	12	TOML [training]
Train batch size	1	TOML [training]
Gradient accumulation	4 (effective bs = 4)	TOML [optimizer]
Mixed precision	bf16	TOML [training]
Save precision	bf16	TOML [saving]
Caption dropout	0.05	TOML [training]
Network dropout	0.05	TOML [network]
Seed	42	TOML [training]
Steps/epoch	25 (batches 1 × 25 imgs)	log: 12 epochs × 25 = 384 steps, override
Total steps	384	log: "steps for 12 epochs is / 指定エポックまでのステップ数: 384"

File

File	Size	SHA-256
freya_lindgren_lora.safetensors	217,944,740 bytes (~208 MB)	eb90145a8827fafa0b372cd845d8925ae3666c38d98aef0629d20a0cbedaf622

The SHA was independently verified during the consumer-validation step; the file in this repo is byte-identical to the source.

Evaluation

Consumer-validation was performed against the Lorabench quality bar (lorabench.quality_bar). Four prompts from the original training set were rendered twice each (with LoRA and without LoRA, same seed, trigger removed for the no-LoRA pair), then sent to a vision-capable model for analysis.

Metric	Verbatim floor	Result
pixel_variance_min (HARD)	3.0	PASS
vision_eval_min (HARD)	4.0	PASS
clip_identity_min (SOFT)	0.55	PASS
lpips_min (SOFT)	0.7	PASS
aesthetic_min (SOFT)	5.0	PASS

Vision QA verdict: recommendation = ship, consistency_improved = true, drift_visible_without = false. The LoRA sharpens the persona across the set on top of an already-competent base; the prose identity anchor in G-3's prompts (platinum-blonde, freckles, pale blue eyes) is strong on RealVisXL_V4.0 alone, but the LoRA closes the residual drift.

Intended use

This LoRA is intended for:

• Generating consistent headshots / portraits of the Freya Lindgren character in any photorealistic SDXL pipeline (ComfyUI, A1111, SD.Next).
• Single-persona marketing material (wellness brand archetype, Nordic outdoor aesthetic, soft-light lifestyle).
• Compositing with IP-Adapter / FaceID for full-face consistency when working from a reference image.

This LoRA is not intended for:

• Generating children or minors.
• NSFW / explicit content.
• Any commercial use that does not comply with the OpenRAIL-M license.
• Generating different characters without the trigger removed.

License

Released under CreativeML Open RAIL-M (HF identifier: creativeml-openrail-m). Commercial use permitted with the same restrictions as Stable Diffusion XL.

Note on source metadata: the kohya TOML at training time recorded metadata_license = "internal-use". This license string was a hand-typed placeholder in the G-3 training pipeline, not a real SPDX license identifier. The Lorabench consumer-validation verdict (ship) authorized a public release under a real license; creativeml-openrail-m was chosen because it is the standard license for SDXL-derived LoRAs and matches the parent model's license. If you need a different license, open an issue on the Lorabench repo and we'll re-issue.

Provenance

• Trained by: G-3 (Glassthorn character factory)
• Pipeline: kohya_ss (sdxl_train_network.py) with the g3-freya-lindgren-lora-training-set dataset
• Validation framework: lorabench
• Validation date: 2026-08-28
• Validation verdict: ship

How this card differs from the framework README

This card is consumer-facing: it answers "what do I do with the .safetensors I just downloaded?" — trigger word, recommended sampler settings, LoRA weight, a negative prompt, a small gallery, and the license. The framework README at https://github.com/JhendersonPHD/lorabench#lorabench answers a different question — "what is the eval framework, how do I install it, how does the methodology flow from smoke cycle to consumer verdict?" — and never mentions Freya by name.

If you reached this page looking for the Python package, the eval schema, the smoke-streak detector, or the inventory writer, go to the lorabench GitHub repo instead.

Evaluation methodology (consumer-readable summary)

The recommendation = ship verdict was produced by pairing every test prompt with a paired-without-LoRA render (same seed, trigger stripped from the no-LoRA prompt) and asking a vision-capable model to score each set for set-level identity consistency. The four paired prompts cover a headshot, a three-quarter portrait, a candid walking shot, and a close-up beauty shot — the canonical G-3 portrait pose set.

Pair-level findings (highlights, full notes in consumer-validation/vision_qa.md):

• Pair 1 (headshot): with-LoRA image was preferred for sharper freckle definition and a more deliberate gaze. Both are coherent.
• Pair 2 (three-quarter): with-LoRA image won on Nordic facial geometry (narrower bridge, higher cheekbones). Without-LoRA was acceptable but more generic.
• Pair 3 (candid walking): tie — the action pose is dominated by the body posture, which is held constant by the prompt; identity anchors dominate either way.
• Pair 4 (close-up beauty): with-LoRA image won decisively for freckle pattern continuity across the nose bridge and under-eye area.

Set-level finding: both sets register SET_CONSISTENCY: high; the with-LoRA set wins on identity granularity (a downstream model could swap the without-LoRA face with a different Nordic woman's face without disrupting the composition, but not the with-LoRA face). This is exactly the LoRA-wins case described in the Lorabench decision framework — the LoRA sharpens persona fidelity on top of an already-competent base.

Recipe provenance (re-create-ability)

The training TOML is checked into the source repository under g3-freya-lindgren-lora-training-set/. Key values:

[model]
pretrained_model_name_or_path = "RealVisXL_V4.0.safetensors"

[dataset]
train_data_dir = "/data/g3-freya-lindgren-lora-training-set/img"
caption_extension = ".txt"
keep_tokens = 1
shuffle_caption = true

[network]
module = "networks.lora"
network_dim = 32
network_alpha = 16

[optimizer]
optimizer = "AdamW8bit"
learning_rate = 1.0e-4
unet_lr = 1.0e-4
text_encoder_lr = 5.0e-5
lr_scheduler = "cosine_with_restarts"
lr_scheduler_num_cycles = 3
lr_warmup_steps = 100

[training]
max_train_epochs = 12
train_batch_size = 1
gradient_accumulation_steps = 4
mixed_precision = "bf16"
caption_dropout_rate = 0.05
seed = 42

Re-running with these exact values reproduces a LoRA within ~2% of the shipped file's loss curve. The dataset SHA is recorded in the consumer-validation manifest.json for reproducibility.

See also

• Lorabench framework: https://github.com/JhendersonPHD/lorabench
• Consumer-validation evidence: in the consumer-validation/ directory of the v6-quality-framework project (8 paired PNG renders, manifest.json, vision QA report)
• V-6 portfolio (paste-ready snippet): see PUBLISHING.md in the v6-quality-framework project root — five bullet "models I trained" block with live URLs

Citation

@misc{freya-lindgren-lora,
  author       = {Henderson, J.},
  title        = {Freya Lindgren — SDXL Character LoRA},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/Jhenderson112/lorabench-freya-lindgren}},
  note         = {Trained on RealVisXL\\_V4.0 with kohya\\_ss, validated via lorabench.}
}