Instructions to use lucentcreative/glycor-coralline-type-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use lucentcreative/glycor-coralline-type-lora with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("lucentcreative/glycor-coralline-type-lora") prompt = "glycor, the lowercase letter s, underwater scene, god rays, dark seafloor background" image = pipe(prompt).images[0] - Inference
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
Glycor โ Coralline Type
A FLUX.1-dev LoRA that renders letters and words in an invented material: coralline glyph โ a white perforated coral-lattice skeleton wrapped in coral and salmon fronds with dusty pink pom-pom seed heads, lit by god-rays over a dark underwater seafloor.
This is the Flux successor to an earlier SDXL experiment; Flux was chosen specifically for its superior letterform and text rendering.
Trigger word
Use the trigger word glycor in the prompt, or at the start of it but paired with the word style.
Usage
glycor style, a single letter "g", underwater scene, god rays, dark seafloor background
glycor style. Text that reads "reef",underwater scene, god rays, dark seafloor
The material vocabulary (coral, lattice, pink, fronds) was deliberately excluded from the training captions, so those words remain free at inference and can be added to intensify the look. Because the base is Flux, whole words render legibly โ not just single letters.
Recommended settings
| Setting | Value |
|---|---|
| Base model | FLUX.1-dev |
| LoRA weight / scale | 0.8 โ 1.0 |
| Inference steps | 28 โ 35 |
| Guidance (CFG) | 3.5 |
| Sampler | FlowMatch / Euler |
| Resolution | 1024 ร 1024 |
Flux uses low guidance (~3.5) and ignores negative prompts โ steer entirely through the positive prompt.
Checkpoints
Trained 2500 steps, saved every 250. Mid-to-late checkpoints (1000โ1750) give the best balance of legibility and material richness.
| Checkpoint | Character |
|---|---|
glycor_v1_000001000 |
Clean letters, lighter material |
glycor_v1_000001250 |
Balanced |
glycor_v1_000001500 |
Richer material, still legible |
glycor_v1 (2500) |
Fullest / most detailed |
Training details
- Tool: Ostris AI-Toolkit (Ostris Cloud), single RTX 4090
- Base model:
black-forest-labs/FLUX.1-dev - LoRA rank: 16 ยท Optimizer: AdamW8Bit ยท LR: 1e-4
- Steps: 2500 ยท batch size: 1
- Resolutions: 512 / 768 / 1024
- Horizontal flip: disabled (letterforms are chiral)
- Dataset: 35 images at 1024ร1024 โ full lowercase aโz, pangram word samples, and material detail shots
License
FLUX.1-dev LoRAs inherit the FLUX.1-dev Non-Commercial License. See the license link above.
Acknowledgements
Built with Ostris AI-Toolkit.
- Downloads last month
- 65
Model tree for lucentcreative/glycor-coralline-type-lora
Base model
black-forest-labs/FLUX.1-dev