📚 LoRA Dataset Forge

Training a LoRA has two miserable parts: wrangling the images into aspect-ratio buckets, and captioning them all. FiLDatasetForge kills both in one run. Feed it a whole image batch (Load Image with a folder of frames), tell it your trigger word, and it buckets the images at the target resolution, writes one LLM caption per frame, and lays the whole thing out as a folder kohya_ss or sd-scripts can eat directly - plus a dataset.toml if you pick the kohya layout. It's the kind of node you reach for when a dataset has been sitting half-built in your head for a month.

How it works

The pipeline is: bucket → crop → caption → write. Buckets are built around base_resolution (1024 for SDXL/Flux, 512–768 for SD 1.5), with bucket_step (64 default) as the side divisor - the standard kohya bucketing math. Each frame is cropped into its bucket with crop_mode center, or entropy if you want the crop pulled toward the most detailed region instead of the middle.

Captioning is where the LLM comes in. With config from the Provider Loader connected, each image gets a caption in caption_mode: natural prose (Flux/SDXL), comma tags (SD 1.5/Pony), hybrid, or none (trigger word only, no API call). You can also bypass the LLM entirely by wiring the captions input with manual captions separated by lines containing only --- - it accepts Optic Scanner output as-is, which is a nice pairing for auditing a small set by hand.

The inputs that decide whether your LoRA generalizes

• trigger_word - the rare token that activates the LoRA, prepended to every caption. Use something like ohwx, not a real word.
• class_token - the class the subject belongs to (woman, car). It follows the trigger in captions and in the kohya folder name (img/<repeats>_<trigger> <class>/).
• base_resolution and caption_mode - match both to your base model.
• dont_caption - the underrated one. List the invariants you never want described ("her face, red hair"); the caption prompt forbids them explicitly.

That last input encodes the single most important rule in LoRA captioning, and the README states it plainly: describe what varies - pose, clothing, background, lighting - and never describe the invariant. The invariant belongs to the trigger word; if your captions describe it, the LoRA learns it as something you'll have to fight at inference time. That rule is why dont_caption exists, and it's the difference between a LoRA that generalizes and one that's welded to your exact dataset.

Outputs and gotchas

Four outputs: preview (the bucketed frames letterboxed onto one square canvas), report (a human-readable summary - read this first), dataset_path (absolute path of what was written), and manifest (per-image bucket, crop, caption and hash). dry_run computes everything and writes nothing - use it, always, the first time.

Two honest warnings. First, the node never upscales: sources smaller than their bucket are still written, but they're flagged in the report's upscaled_count. Small images in a dataset quietly hurt training, so run stragglers through FiL Upscaler Simple first. Second, multi-subject attribution is the known weak spot of every captioning VLM - if your set has multiple people doing different things, audit those captions by hand rather than trusting the LLM. Manual captioning is still the gold standard for small sets, and the captions input makes that easy.

Install is pack-standard: Manager search FiL_Design_ImageMind, or clone + pip install -r requirements.txt + restart. No captioning requires no config at all; LLM captions need a provider.