Potiron/lavender-flux

English lavender (Lavandula angustifolia) — FLUX.1 [dev] LoKr + LoRA

A LoKr (LyCORIS) and a LoRA for FLUX.1 [dev] that generate close-ups of English lavender (Lavandula angustifolia), with the growth stage set by the prompt. Part of a personal series of botanical models (climbing plants, garden plants, houseplants) aimed at accurate plant details for game assets and illustration.

The LoKr learned the real foliage habit of lavender: woody, branching stems with short silvery leaves and old dried flower stems. FLUX.1 [dev] alone adds flower spikes even when asked for a plant in leaf, without flowers; the LoRA respects the stage but draws a grass-like tuft.

Which file to choose

File	Type	Size	Epoch
lavender_lvnd_zc_lokr.safetensors	LoKr (LyCORIS)	370 MB	4 of 4 (1,804 steps)	Best overall: woody branching foliage, natural flower spikes
lavender_lvnd_zc_lora.safetensors	LoRA, rank 16	153 MB	4 of 4 (1,804 steps)	Flowers as good as the LoKr; foliage more tuft-like.

Both were trained on the same images and captions. Use the LoRA if your tool only accepts standard Flux LoRAs (LoKr needs LyCORIS support; ComfyUI and Forge load both). Load only one of the two at a time: loaded together, the LoKr cancels the LoRA in some inference scripts.

Trigger and growth stages

Start the prompt with the trigger lvnd_zc, then the plant's common and scientific names and the stage, as in the captions:

Stage	Prompt opening
bloom	lvnd_zc, English lavender (Lavandula angustifolia) in bloom
foliage	lvnd_zc, English lavender (Lavandula angustifolia) in leaf, without flowers

Then add details from the controlled vocabulary used in the captions, then Background: …, light, framing and No people.:

• flower colour: purple, deep purple, bright violet, blue-violet, pale lilac, lavender-blue
• flower state: open flowers, tight flower buds, open flowers and buds, fading flowers with brown calyxes
• number: A single flower spike, A few flower spikes, Many flower spikes; insects: with a bumblebee, a bee, a small butterfly
• foliage: silvery-grey, grey-green, fresh green narrow leaves; woody stems, old dried flower stems, a rounded mound
• background: blurred garden, bare soil, grass, dry grass, gravel, stones, stone wall, moss, mulch, fallen leaves, dark background…
• light and framing: Sunny, Overcast, Shade; macro close-up, close-up, medium shot

Recommended settings

• FLUX.1 [dev] (fp8 is fine), 768 × 768 (training resolution), 28 steps, euler / simple, CFG 1 and FluxGuidance 3.5.
• Weight 1.0; 0.7–0.8 if the image copies the training photos too closely.
• ComfyUI: standard LoraLoaderModelOnly node; a workflow is in comfyui/.

Examples

Samples generated at the end of training (seed 42, 768 px).

_{LoKr · bloom · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in bloom, with purple flower spikes and open flowers. A single flower spike, with a bumblebee. Background: blurred garden. Sunny, macro close-up. No people.}

_{LoKr · bloom · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in bloom, with deep purple flower spikes and tight flower buds. Many flower spikes. Background: gravel. Overcast, close-up. No people.}

_{LoKr · foliage · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in leaf, without flowers, with silvery-grey narrow leaves and woody stems. Background: fallen leaves. Overcast, close-up. No people.}

_{LoKr · foliage · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in leaf, without flowers, with grey-green narrow leaves, old dried flower stems and a rounded mound. Background: bare soil. Overcast, medium shot. No people.}

_{LoRA · bloom · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in bloom, with purple flower spikes and open flowers. A single flower spike, with a bumblebee. Background: blurred garden. Sunny, macro close-up. No people.}

_{LoRA · bloom · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in bloom, with deep purple flower spikes and tight flower buds. Many flower spikes. Background: gravel. Overcast, close-up. No people.}

_{LoRA · foliage · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in leaf, without flowers, with silvery-grey narrow leaves and woody stems. Background: fallen leaves. Overcast, close-up. No people.}

_{LoRA · foliage · seed 42 · 768 px · lvnd_zc, English lavender (Lavandula angustifolia) in leaf, without flowers, with grey-green narrow leaves, old dried flower stems and a rounded mound. Background: bare soil. Overcast, medium shot. No people.}

Comparisons

FLUX.1 [dev] alone vs + LoRA vs + LoKr

Same prompts and seeds (42 and 7), 768 px. Rows: single spike with a bumblebee, buds on gravel, foliage with woody stems, rounded mound with dried stems (row labels in French). Flux alone puts flowers on the foliage prompts; the LoKr draws the woody, branching habit.

Training

Base	FLUX.1 [dev], fp8, transformer only (text encoders frozen)
Network	LoKr: LyCORIS, factor=8, preset=full, full w2 (dim 100000, alpha 1). LoRA: kohya networks.lora_flux, rank 16, alpha 16
Data	451 close-up photos (278 in bloom, 173 in leaf), 373 different plants; captions written per photo in sentences with a closed vocabulary
Resolution	768 px, aspect-ratio buckets
Optimizer	AdamW8bit, LR 1e-4, cosine, batch 1
Flux specifics	model_prediction_type=raw, timestep_sampling=shift, discrete_flow_shift=3.1582, guidance 1.0
Tools, hardware	kohya sd-scripts + LyCORIS, RTX 5080 Laptop 16 GB

Limitations

• Only Lavandula angustifolia: French/Spanish lavender (L. stoechas), L. dentata and white cultivars were removed from the dataset.
• Trained on close-ups (macro to medium shot): lavender fields and wide landscapes are out of scope.

Data and license

• Trained on 451 photos from iNaturalist and Wikimedia Commons under CC0, CC-BY or public domain, checked one by one. CC-BY authors are credited in CREDITS.md. The training dataset is not redistributed.
• These weights are a derivative of FLUX.1 [dev] and are released under the FLUX.1 [dev] Non-Commercial License: non-commercial use only.