darkroom-noir-zimage

DARKROOM Noir

A dark tech-noir style LoRA for Z-Image Turbo. Deep navy, disciplined blue light, premium restraint — applied to any prompt you can type.

🌐 Showcase & interactive training scrubber: https://refxfrank.github.io/darkroom-noir/

All samples: Z-Image Turbo + DARKROOM Noir @ 0.8, 8 steps, unretouched.

Usage

Start your prompt with the trigger: drkrm style,

ComfyUI recipe:

Works with any Z-Image Turbo quantization, including the official int8-ConvRot repack (6.2 GB — runs on 16 GB cards with lots of headroom).

What it does

This is a style adapter, not a subject adapter: your subject, its palette. The training set contains only abstract wallpapers, night scenes and dark macro shots — yet portraits, objects and scenes it has never encountered come out in the same aesthetic: deep navy-to-black gradients, a single disciplined blue accent, near-zero color pollution.

Training

• Trainer: ai-toolkit (Ostris), headless, on a single RTX 5080 (16 GB)
• Method: LoRA rank 32 / alpha 32, bf16, 3,000 steps, flow-matching with weighted timesteps, trained against Z-Image Turbo through the de-distillation training adapter so the result samples cleanly on stock Turbo at 8 steps
• Dataset: 30 images, 100% self-generated in [the author's local DARKROOM studio] and hand-curated against a strict palette/restraint rubric, captioned locally with WD14 (EVA02-Large v3). Zero scraped or third-party content.
• Stack cost: $0 — every model, tool and byte of training data in the chain is free and Apache-2.0

License

Apache-2.0 — the same license as the base model, the training adapter, and the dataset provenance chain. Free for any use, including commercial. Attribution appreciated, not required.

---

Developed end-to-end in DARKROOM, a fully-local, zero-cloud, zero-telemetry image studio built on ComfyUI.