ArtificialGuyBR

Home / Blog / Training LoRAs

Crochet Amigurumi LoRA for Z-Image-Turbo: Consistent Yarn-Style Generation

Generate consistent crochet amigurumi images with this LoRA for Z-Image-Turbo. Trigger words, settings, and example outputs.

1 sources cited Training LoRAs

On this page

  1. What this LoRA does
  2. How to use
  3. Recommended settings
  4. Trigger words
  5. License
  6. Common errors and fixes
  7. FAQ
  8. Sources

Crochet Amigurumi LoRA for Z-Image-Turbo

What this LoRA does

This LoRA adapts the Z-Image-Turbo text-to-image model to specialize in crochet amigurumi aesthetics. Amigurumi is the Japanese art of crocheting small, stuffed yarn creatures — think rounded forms, visible stitch texture, and the soft, handmade quality of yarn toys. The adapter was trained on a curated dataset of amigurumi photography with GPU compute sponsored by Redmond.AI.

Example amigurumi generation showing a crocheted character with yarn texture and stitch detail

The result is a lightweight adapter (safetensors format) that adds strong stylistic control without requiring a full model finetune. Outputs consistently show yarn fiber texture, stitch definition, and the characteristic proportions of amigurumi figures.

How to use

  1. Load Z-Image-Turbo as your base checkpoint
  2. Add a Load LoRA node pointing to the downloaded .safetensors file
  3. Set LoRA strength around 0.8–1.0 (start at 1.0)
  4. Include the trigger phrase Amigurumi, Crochet. at the start of your prompt

diffusers (Python)

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.float16
).to("cuda")

pipe.load_lora_weights("artificialguybr/CROCHET-AMIGURUMI-REDMOND-ZIMAGE")

prompt = "Amigurumi, Crochet. A cute crocheted bear sitting on a shelf, soft yarn texture, warm lighting"
image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]

Z-Image-Turbo is a distilled/fast model — use low step counts (4–8) and guidance_scale near 0.

Second example showing amigurumi animals with consistent stitch quality

Parameter | Value

|---|---

Steps | 4–8 Guidance scale | 0.0–1.0 LoRA weight | 0.8–1.0 Resolution | 1024×1024 (native) Sampler | Euler a / DPM++ 2M

Trigger words

The model card specifies a single compound trigger:

Amigurumi, Crochet.

Place this at the beginning of your prompt. The period matters — the training data used this exact formatting.

License

Apache 2.0 — permissive, allows commercial use, modification, and distribution with attribution. See the HF model card for the full license text.

Common errors and fixes

Error | Cause | Fix

|---|---|---

Style not activating | Missing trigger phrase | Prepend Amigurumi, Crochet. to every prompt Overcooked / burnt look | LoRA weight too high | Reduce to 0.7–0.8 Blurry / low detail | Too many steps on Z-Image-Turbo | Use 4–8 steps, guidance ~0 Wrong base model | Using SDXL/Flux checkpoint | Must use Tongyi-MAI/Z-Image-Turbo

FAQ

Q: Does this work with SDXL or Flux? A: No. This LoRA targets Z-Image-Turbo specifically. The architecture and latent space differ.

Q: Can I merge this LoRA into the base model? A: Yes — use merge_lora in ComfyUI or the diffusers pipe.merge_lora() method for a standalone checkpoint.

Q: What's the difference between this and the FluxKlein amigurumi LoRA? A: Different base models. This targets Z-Image-Turbo (fast, 4-step); the FluxKlein version targets Flux.1-Klein. Choose based on which base model you prefer.

Q: Are the example images cherry-picked? A: The model card displays 15 widget samples (images 001–015). The three shown here are representative of the consistent style.

Sources