flfwrld — Fluffy 3D Style LoRA
A style LoRA producing soft, felted, candy-colored 3D illustrations: round plush
characters, fuzzy pom-pom landscapes, and pastel palettes.
- Base model: black-forest-labs/FLUX.2-klein-base-4B
- Trigger word:
flfwrld
- Training: AI-Toolkit on Ostris Cloud · rank 32 · LR 1e-4 · 3000 steps · batch 1
- Dataset: 35 curated images, captioned content-only (no style words) so the
style binds to the trigger word
- Created for: LABASAD master's assignment "Create a LoRA"
Checkpoints
| File | Step | Character |
|---|
flfwrld_style_v1_000000500.safetensors | 500 | Palette converged, texture weak |
flfwrld_style_v1_000001000.safetensors | 1000 | Texture arriving, balanced |
flfwrld_style_v1_000001250.safetensors | 1250 | Best prompt adherence with plain prompts |
flfwrld_style_v1.safetensors | 3000 | Strongest felt texture — recommended with texture words in the prompt |
Recommended inference
Tested on fal.ai (fal-ai/flux-2/klein/4b/base/lora):
steps 28–40 · guidance 4 · LoRA scale 1.0 · no acceleration.
Note: the undistilled base needs 25+ inference steps — few-step endpoints
produce hazy, unconverged images.
Prompt template:
flfwrld style, [subject + action], fuzzy felted wool texture, soft plush surfaces, [background], [lighting], [composition]
Because style words were excluded from training captions, adding them at
inference ("fuzzy felted wool texture") stacks with the trigger and
strengthens the texture. With these words present, the 3000 checkpoint gives
the best combination of texture and prompt adherence.
