beenga-curl-v1

A LoRA adapter for FLUX.2 [klein] 4B that sharpens soft and salon-curl hair geometry, trained as part of Beenga Image.

Read the limitations before using this. The adapter works, and it leaks. Both halves are documented below, because the leak is the reason it is opt-in rather than default in the product it was built for.

Adapter	LoRA, rank 32, alpha 32
Base model	black-forest-labs/FLUX.2-klein-4B (Apache 2.0)
Checkpoint	step 500 of a 1500-step run — see Which checkpoint below
File	beenga_curl_v1.safetensors, 160 tensors, 92,426,528 bytes
SHA-256	bea8e082a3ed30dd63d37a217a726a9e2f60422cfba2e8de4c453a601ecef6b4
Trained with	ai-toolkit 0.12.23

Usage

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16,
).to("cuda")

pipe.load_lora_weights(
    "beenga8/beenga-curl-v1",
    weight_name="beenga_curl_v1.safetensors",
    adapter_name="curl",
)

image = pipe(
    "a young indian woman with soft salon curls, rooftop in Delhi",
    num_inference_steps=4,
).images[0]

Always pass weight_name explicitly. If you run with HF_HUB_OFFLINE=1 or local_files_only=True — which you may well need to, on serverless workers that cannot reach the Hub during setup — the loader will not guess a filename and raises ValueError: When using the offline mode, you must specify a 'weight_name' before it ever looks at the path. A valid local file fails with a message that does not say so. This cost us a shipped feature that failed on 100% of calls; the explicit form is correct either way.

To unload:

pipe.unload_lora_weights()

What it does

Asked for curls, it renders better curl geometry than the base model. Explicit control in the other direction survives intact: ask for pin-straight hair and it renders pin-straight hair.

What it does wrong

This is the important section.

• Unspecified hair drifts curly. If the prompt says nothing about hair, the adapter pushes toward curls anyway. It changes a default you did not ask it to change.
• The training set's look bleeds into unrelated scenes. Plainer backgrounds, more ordinary faces, deeper complexions, flatter and more neutral expressions — applied regardless of what the prompt asked for.

The cause is the dataset, not the training run. All 200 captions came from a single template and a single generator, with no contrast examples — nothing straight, tight, coily or glamorous. The adapter had no way to learn that curl geometry is separable from everything else in the frame, so it learned "curls" and "this visual style" as one thing.

If you use this, use it deliberately, on prompts where curls are the subject. In Beenga Image it is exposed as an opt-in flag that is off by default, for exactly this reason.

Which checkpoint

Checkpoints were saved at steps 500, 1000 and 1500. Step 500 is published here because it is the best of the three; step 1500 is visibly overtrained.

Stated plainly because it matters to anyone comparing: the Beenga Image production image currently serves the step-1500 file, which is a packaging mistake — the final checkpoint was saved under the plain output name and that is the name the predictor loads. It is recorded in the project's model card and will be corrected on the next rebuild. The file in this repository is step 500, verified from its own training_info metadata ({"step": 500, "epoch": 0}) and by the SHA-256 above.

Training

Steps	500 (published) of 1500 run
Optimiser	adamw8bit, lr 1e-4
Scheduler	flowmatch
Precision	bf16, quantised
Resolutions	512 / 768 / 1024
Hardware	1× NVIDIA A40 48GB, ~35 min for the full 1500-step run

Training data

Source	Share	Licence
Z-Image Turbo generations	100%	Apache 2.0, no output restriction
Real photographs	0%	—

200 images, entirely synthetic. No scraped data, no stock imagery, no images of identifiable real people, no user-contributed photographs. Generation recipes are in datasets/recipes.mjs.

Evaluation

Assessed by human judgement against the Beenga Image benchmark suites, with fixed seeds so differences are attributable to the adapter rather than sampling noise. Scoring for hair geometry is manual and therefore subjective and not reproducible — a real limitation, and the same one that applies to every axis of that project except complexion, which is measured by script.

Licence

Apache 2.0, matching the base model. Use of FLUX.2 [klein] 4B is also subject to Black Forest Labs' Out-of-Scope Use policy, which binds conduct separately from the copyright licence.

Check the variant before substituting. As of 2026-08-16, Black Forest Labs publishes the 4B models under Apache 2.0 and the 9B models under the FLUX Non-Commercial License v2.1. The names differ by two characters and both live in the same organisation.

This adapter introduces no safety or moderation capability of its own.

Beenga™ is a trademark of Beenga. Apache 2.0 grants no trademark rights.