dltwrkn_1920

FLUX.1-dev LoRA. Trigger token: dltwrkn

Training configuration

Base model	flux1-dev.safetensors (BFL checkpoint)
Dataset	30 images, 6 repeats, 20 epochs
Steps	3600 (180 per epoch)
Per-image exposure	120
Resolution	1920 × 2880, enable_bucket = false
Network	dim 32 / alpha 16, networks.lora_flux
Learning rate	5e-4, cosine, 1 cycle
Text encoder LR	0
Optimizer	adamw8bit
Batch size	1
Precision	bf16 (save fp16)
Noise offset	0.05
Timestep sampling	shift, model_prediction_type raw, guidance 1.0
Attention	sdpa, gradient checkpointing on
Seed	42
Trainer	sd-scripts @ b8d1eb067eba32bb105984678b97f05b11452940
Init	from scratch — no resumed weights

Bucketing was disabled rather than configured: all 30 source files are exactly 1920 × 2880, so the bucketing mechanism had nothing to resolve. A mis-sized file therefore fails loudly instead of being silently rebucketed.

Checkpoints published: epochs 12, 15, 18, 20 (dltwrkn_1920.safetensors is epoch 20).

Caption scheme

• dltwrkn is the first token of every caption. keep_tokens = 1, shuffle_caption = false.
• Every element that varies image to image is named.
• Framing is named on close crops.
• Material is never named.

Constants left deliberately unnamed: violet mass, orange contours, mesh, particles, black void.

Two phrases are verbatim in the corpus and must match at inference:

Phrase	Instances	Expected binding
seen through	7	supported
bridges of light	2	partial

Corpus tag count: 91 distinct tags across 30 captions.

Material model

Volume is carried by contour behaviour — weave compressing at a turning edge, ribbons riding curvature, foreshortening toward a silhouette — not by fill. A surface fails when the lattice runs straight across it and does not compress at the silhouette.

Three material classes, applied by rule but not consistently across the corpus:

• bare skin — surface with weave, no interior
• cloth — thin sheet
• props / objects — holographic shell

Opaque vs. translucent has no corpus rule; the same figure can resolve both ways within one frame. A global caption cannot route this regionally — per-region control would require masking on the conditioning.

Predecessor and open question

The previous LoRA (dltwrkn_1536) was trained on 2 caption tags total and showed prompt-independent residue: unrequested hands, scrollwork, tables, grid backgrounds. Its recorded per-image exposure was also double what was intended (duplicate files in a .ipynb_checkpoints directory were trained alongside the originals), placing it near the memorisation regime.

This run carries 91 tags, a clean file list, and a higher training resolution. Whether the caption scheme suppresses the residue is the open question.

Attribution caveat: corpus, caption scheme, and training resolution all changed together relative to the predecessor. A clean result is therefore not attributable to the caption scheme alone. The residue features (hand, scrollwork, table, grid) are the more diagnostic signal, since resolution has no plausible mechanism for suppressing them; material quality is confounded.

Usage notes

Recommended strength 1.00. A 1.00 / 1.12 / 1.50 sweep on the predecessor moved the weave finer and thinner rather than toward the target, and composition bleed appeared at 1.12 — strength is not the dial for material quality.