Generation pipeline scripts and LoRA weights for the Coalition's FLUX.1-dev image generation stack.
Repository Structure
scripts/ # 29 generation pipeline scripts
loras/
vera-likeness/ # Vera character likeness LoRAs (v1, v3, v4)
kintsugi-texture/ # Kintsugi gold-repair texture style LoRAs (v1, v2)
thomas-likeness/ # Thomas character likeness LoRA (v1)
Scripts
Generation pipeline scripts from the vera-triple-stack workspace. These drive FLUX.1-dev inference with single or stacked LoRAs for various visual styles and compositions.
Key scripts:
gen_v2.py through gen_v5.py -- base generation pipeline iterations
gen_cached_identity.py / precompute_identity.py -- identity embedding caching for faster generation
All LoRAs are trained on FLUX.1-dev with LoRA rank 16, trained on Apple Silicon (MPS).
Vera Likeness (loras/vera-likeness/)
Character likeness LoRA for Vera. Trigger token: vera.
File
Version
Notes
vera_likeness_v1.safetensors
v1
Initial training, 750 steps
vera_likeness_v3.safetensors
v3
Updated prompts with ceramic/statuesque aesthetic, 1250 steps
vera_likeness_v4.safetensors
v4
Fine-tuned from v3, +750 steps. Best version.
config_v1.yaml
v1
Training configuration
config_v3.yaml
v3
Training configuration
config_v4.yaml
v4
Training configuration
v2 was an incomplete training run and is not included.
Kintsugi Texture (loras/kintsugi-texture/)
Style LoRA for kintsugi (gold-repair) texture effects. Applies golden crack/seam patterns inspired by the Japanese art of repairing broken pottery with gold.
File
Version
Notes
kintsugi_texture_v1.safetensors
v1
Initial texture training
kintsugi_texture_v2.safetensors
v2
Refined texture, 300 steps
config_v1.yaml
v1
Training configuration
config_v2.yaml
v2
Training configuration
Thomas Likeness (loras/thomas-likeness/)
Character likeness LoRA for Thomas. Trained August 2026.
File
Version
Notes
thomas_likeness_v1.safetensors
v1
1250 steps
config_v1.yaml
v1
Training configuration
Usage
These LoRAs are designed for use with FLUX.1-dev via diffusers. See the generation scripts for examples of single-LoRA and stacked multi-LoRA inference.
Basic single-LoRA usage:
python
1from diffusers import FluxPipeline
2import torch
34pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)5pipe.load_lora_weights("LiberationLabs/image-toolbench", weight_name="loras/vera-likeness/vera_likeness_v4.safetensors")6pipe.to("cuda")# or "mps" for Apple Silicon78image = pipe("portrait of vera, ceramic aesthetic, golden light", num_inference_steps=30).images[0]
