kirusanth08/flux-dev-sty1ref-lora

sty1ref — FLUX.1-dev style LoRA (v1)

A stylised portrait-illustration LoRA: hard-edged planar shading, flat colour blocking across the face, visible geometric facets in skin and hair, muted palette.

Trigger word: sty1ref — put it at the front of the prompt.

Recommended settings

Setting	Value
Checkpoint	sty1ref-step00002000.safetensors
LoRA strength	1.4 – 1.6 (not 1.0 — see below)
Base	flux1-dev (fp8_e4m3fn is fine)
Sampler / scheduler	euler / beta, 28 steps
FluxGuidance	3.0
CFG	1.0 (Flux dev is CFG-distilled; negatives do nothing)

Read this before using it

Strength 1.0 is too weak. This LoRA is undertrained at unit strength — a prompt rendered at 1.0 comes back looking like base Flux with a light stylisation pass. The faceted planar shading only appears clearly from about 1.4 upward. 1.5 is the sweet spot; 2.0 works but muddies the midtones.

It is portrait-biased. All 12 training images are head-and-shoulders portraits, so the trigger has only ever co-occurred with faces. Consequences:

• Portraits of unseen subjects: works well at 1.5.
• Scenes without people (streets, landscapes, objects): the output becomes painterly but does not pick up the hard faceted planes. At 1.0 it is essentially unstyled.

If you need this texture on arbitrary scenes, retrain with non-portrait references in the same style — full figures, architecture, objects, landscapes. That is the fix; no strength value substitutes for it.

Checkpoints

File	Steps	Notes
sty1ref-step00000400.safetensors	400	barely stylised
sty1ref-step00000800.safetensors	800	faint
sty1ref-step00001200.safetensors	1200	usable at 1.5
sty1ref-step00001600.safetensors	1600	close second
sty1ref-step00002000.safetensors	2000	recommended

samples/ holds two renders per checkpoint at a fixed seed (42) — one portrait of an unseen subject, one people-free scene — plus the comparison grids.

Training recipe

kohya-ss sd-scripts (sd3 branch), flux_train_network.py, on one RTX 5090 (32 GB), ~1 hour for 2000 steps.

--network_module networks.lora_flux --network_dim 24 --network_alpha 24
--network_train_unet_only
--optimizer_type adamw8bit --learning_rate 1e-4
--lr_scheduler constant_with_warmup --lr_warmup_steps 40
--max_train_steps 2000 --save_every_n_steps 400
--gradient_checkpointing --mixed_precision bf16 --fp8_base --sdpa --highvram
--timestep_sampling shift --discrete_flow_shift 3.1582
--model_prediction_type raw --guidance_scale 1.0 --loss_type l2

Dataset: 12 images, aspect-ratio bucketing at 1024 base with bucket_no_upscale, keep_tokens = 1. Captions describe content only (subject, clothing, framing, background) so that everything constant across the set collapses onto the trigger token rather than scattering across style adjectives.

Known cause of the weak transfer: 12 images at lr 1e-4 for 2000 steps undercooks this style. A v2 should use more images and/or lr 2e-4.

Dataset

kirusanth08/sty1ref-dataset (private) — the 12 cleaned images and their caption files.

Usage (ComfyUI)

workflow/workflow_api.json is a working API-format graph at the recommended settings. workflow/workflow.json drags onto the canvas.

Licence

Inherits the FLUX.1-dev non-commercial licence. The training references were collected from the web and are not owned by the author of this LoRA; treat outputs accordingly.