Chunte
/

huggy-style-v1-lora

@@ -1,121 +1,128 @@
----
-base_model: black-forest-labs/FLUX.1-dev
-library_name: diffusers
-license: other
-license_name: flux-1-dev-non-commercial-license
-license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
-inference: true
-tags:
-- flux
-- flux-diffusers
-- lora
-- diffusers
-- text-to-image
-- character
-- mascot
-- dreambooth
-widget:
-- text: "a huggy_style_v1 mascot wearing a pirate hat, waving, happy"
-- text: "a huggy_style_v1 mascot wearing sunglasses, thumbs up"
-- text: "a huggy_style_v1 mascot holding a heart, smiling"
-- text: "a huggy_style_v1 mascot wearing a chef hat, open hands raised"
----
-# Huggy Style v1 - FLUX DreamBooth LoRA
-A LoRA adapter for [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) trained with DreamBooth to generate **Huggy** - the HuggingFace mascot character.
-## Character Description
-Huggy is a yellow circular character with:
-- Round body with no arms, legs, or feet
-- Two floating hands
-- Orange outlines (no dark black outlines)
-- Clean flat vector art style with edge shadows
-- Expressive face with various emotions
-## Sample Outputs
-![img_0](./image_0.png)
-![img_1](./image_1.png)
-![img_2](./image_2.png)
-![img_3](./image_3.png)
-## Usage
-Use the trigger word **`huggy_style_v1`** in your prompts.
-```python
-import torch
-from diffusers import FluxPipeline
-pipe = FluxPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-dev",
-    torch_dtype=torch.bfloat16,
-)
-pipe.load_lora_weights("Chunte/huggy-style-v1-lora")
-pipe.to("cuda")
-image = pipe(
-    "a huggy_style_v1 mascot wearing a pirate hat, waving, happy",
-    num_inference_steps=28,
-    guidance_scale=3.5,
-).images[0]
-image.save("huggy_pirate.png")
-```
-## Prompt Tips
-- Always include `huggy_style_v1` as the trigger word
-- Describe what varies: outfit, pose, expression, props
-- Keep prompts simple and descriptive
-- Example: `a huggy_style_v1 mascot wearing a santa hat, holding a gift, happy`
-## Training Details
-| Parameter | Value |
-|---|---|
-| Base model | FLUX.1-dev |
-| Method | DreamBooth LoRA |
-| Training images | 72 |
-| Resolution | 768 |
-| LoRA rank | 32 |
-| Learning rate | 1e-4 |
-| LR scheduler | constant (100 warmup steps) |
-| Training steps | 2000 |
-| Batch size | 1 (x4 gradient accumulation) |
-| Mixed precision | bf16 |
-| Guidance scale | 1.0 |
-| Seed | 42 |
-| Hardware | NVIDIA L40S (48GB) |
-| Training time | 2h 45min |
-| Final loss | 0.021 |
-### Loss Curve
-| Step | Loss |
-|---|---|
-| 1 | 0.050 |
-| 100 | 0.267 |
-| 500 | 0.078 |
-| 1000 | 0.135 |
-| 1500 | 0.039 |
-| 2000 | 0.021 |
-## Dataset
-Trained on 72 hand-curated images of the Huggy character in various outfits, poses, and expressions. Each image has a per-image natural language caption describing only what varies (outfit, pose, expression) while the trigger word captures the character identity and art style.
-Dataset: [Chunte/huggy_for_training](https://huggingface.co/datasets/Chunte/huggy_for_training)
-## Checkpoints
-Intermediate checkpoints are available at steps 500, 1000, 1500, and 2000. To load a specific checkpoint:
-```python
-pipe.load_lora_weights("Chunte/huggy-style-v1-lora", subfolder="checkpoint-1000")
-```
-## License
-This LoRA inherits the [FLUX.1-dev non-commercial license](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md).

+---
+base_model: black-forest-labs/FLUX.1-dev
+library_name: diffusers
+license: other
+license_name: flux-1-dev-non-commercial-license
+license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
+inference: true
+tags:
+- flux
+- flux-diffusers
+- lora
+- text-to-image
+- diffusers-training
+- dreambooth
+- dreambooth-lora
+- character
+- template:sd-lora
+widget:
+- text: "a huggy_style_v1 mascot wearing a pirate hat, waving, happy"
+  output:
+    url: image_0.png
+- text: "a huggy_style_v1 mascot wearing a chef hat, holding a pizza"
+  output:
+    url: image_1.png
+- text: "a huggy_style_v1 mascot in a spacesuit, floating in space"
+  output:
+    url: image_2.png
+- text: "a huggy_style_v1 mascot sitting on a stack of books, reading"
+  output:
+    url: image_3.png
+---
+# Huggy Style v1 - FLUX DreamBooth LoRA
+A LoRA adapter for [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) trained with DreamBooth to generate **Huggy** — the HuggingFace mascot character.
+## Character Description
+Huggy is a **yellow circular character** with:
+- Round body (no arms, legs, or feet)
+- Two floating hands
+- Orange outlines (no dark black outlines)
+- Clean flat vector art style with edge shadows
+- Expressive face with various emotions
+## Trigger Word
+Use **`huggy_style_v1`** in your prompts to activate the character.
+## Usage
+```python
+import torch
+from diffusers import FluxPipeline
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    torch_dtype=torch.bfloat16,
+)
+pipe.enable_model_cpu_offload()
+# Load LoRA
+pipe.load_lora_weights("Chunte/huggy-style-v1-lora")
+image = pipe(
+    prompt="a huggy_style_v1 mascot wearing a pirate hat, waving, happy",
+    num_inference_steps=28,
+    guidance_scale=3.5,
+    width=768,
+    height=768,
+    generator=torch.Generator("cpu").manual_seed(42),
+).images[0]
+image.save("huggy.png")
+```
+## Prompt Tips
+- Always include `huggy_style_v1` as the trigger word
+- Describe **what varies** — costumes, poses, expressions, props
+- Don't describe the character's base appearance (yellow, circular, etc.) — the LoRA already knows this
+- Example: `a huggy_style_v1 mascot wearing a santa hat, holding a gift, smiling`
+## Checkpoints
+Multiple checkpoints are available if the final weights are overfitting:
+| Checkpoint | Use Case |
+|-----------|----------|
+| `checkpoint-500` | Early training — more creative, less accurate character |
+| `checkpoint-1000` | Moderate — good balance for some use cases |
+| `checkpoint-1500` | Strong character identity with good generalization |
+| **final (default)** | **Strongest character identity** (2000 steps) |
+Load a specific checkpoint:
+```python
+pipe.load_lora_weights("Chunte/huggy-style-v1-lora", subfolder="checkpoint-1000")
+```
+## Training Details
+| Parameter | Value |
+|-----------|-------|
+| Base model | FLUX.1-dev |
+| Method | DreamBooth LoRA |
+| Training script | `train_dreambooth_lora_flux.py` (diffusers v0.37.0) |
+| Dataset | 72 hand-captioned images (1024x1024, white background) |
+| Resolution | 768 |
+| LoRA rank | 32 |
+| Learning rate | 1e-4 (constant scheduler) |
+| Warmup steps | 100 |
+| Training steps | 2000 |
+| Batch size | 1 (gradient accumulation: 4, effective batch: 4) |
+| Mixed precision | bf16 |
+| Guidance scale | 1 (recommended for FLUX training) |
+| Gradient checkpointing | Enabled |
+| Hardware | NVIDIA L40S (48GB VRAM) |
+| Final loss | 0.021 |
+## Sample Images
+![img_0](./image_0.png)
+![img_1](./image_1.png)
+![img_2](./image_2.png)
+![img_3](./image_3.png)
+## License
+This LoRA adapter inherits the [FLUX.1-dev Non-Commercial License](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md).