Dapple — Local AI Image Generator

A single-file, interactive terminal tool for generating images with all major FLUX models via HuggingFace Diffusers. Works entirely on your local machine — no cloud API, no subscription, no external server.

Quick Start

# 1. Install PyTorch with CUDA first:
#    https://pytorch.org/get-started/locally/

# 2. Install dependencies:
pip install -r requirements.txt

# 3. (Optional) Create .env file with your HF token:
#    echo HF_TOKEN=hf_your_token_here > .env

# 4. Run the generator:
python dapple.py

Available Models

Key	Model	Params	Speed	License	Token?
klein-4b	FLUX.2 Klein 4B	4B	Fastest	Apache 2.0	No
klein-9b	FLUX.2 Klein 9B	9B	Fast	Non-commercial	Yes
schnell	FLUX.1 Schnell	12B	Moderate	Apache 2.0	No
dev	FLUX.1 Dev	12B	Slow	Non-commercial	Yes
flux2-dev	FLUX.2 Dev	32B	Slowest	Non-commercial	Yes

Run python dapple.py --list to see full model details.

How to Use

Interactive Mode (default)

1. The script asks "Which model would you like to load?" — pick by number or key.
2. The model loads (first run downloads ~10–45 GB, cached afterward).
3. Type your prompt (required) — describe the image you want.
4. Type a negative prompt (optional, press Enter to skip) — describe what to avoid.
5. The image is generated and saved to Generated_Images_YYYY-MM-DD/.
6. Repeat! Type quit, q, or exit to stop.

Configuration Mode

Edit generator_config.json to pre-set values. When any config is active, the script:

• Skips model selection (if model is set)
• Uses the configured prompt (if default_prompt is set)
• Uses the configured negative prompt (if default_negative_prompt is set)
• Shows a notice that config mode is active

{
    "model": "klein-4b",
    "default_prompt": "A serene mountain lake at sunset, photorealistic",
    "default_negative_prompt": "blurry, low quality, distorted",
    "cache_dir": "D:/AI_Models",
    "steps": 4,
    "use_nf4": true
}

Leave any field as "" or 0 to use the interactive flow for that step.

HuggingFace Token

Some models require a free HuggingFace token:

1. Accept the license at the model's HF page (linked in the model list)
2. Create a token at https://huggingface.co/settings/tokens
3. Create a .env file next to dapple.py:

   HF_TOKEN=hf_your_token_here
   

Models that don't need a token: klein-4b, schnell

Output

Images are saved to a folder named Generated_Images_YYYY-MM-DD/ (e.g., Generated_Images_2026-06-27/). Each filename includes the timestamp and a truncated version of your prompt.

Configuration Reference (generator_config.json)

Field	Type	Default	Description
model	string	""	Pre-select model. One of: schnell, dev, klein-4b, klein-9b, flux2-dev
default_prompt	string	""	Default prompt. If set, prompt step is skipped
default_negative_prompt	string	""	Default negative prompt. If set, negative step is skipped
cache_dir	string	""	Model download location. Empty = ~/.cache/huggingface
steps	int	0	Inference steps. 0 = use model default
guidance	float	0.0	Guidance scale. 0.0 = use model default
width	int	0	Image width. 0 = use model default
height	int	0	Image height. 0 = use model default
seed	int	-1	Random seed. -1 = random each time
use_nf4	bool	true	4-bit NF4 quantization (~¼ VRAM)
use_cpu_offload	string	"auto"	"auto", "always", or "never"

VRAM Requirements (approximate)

Model	Params	NF4	Full
FLUX.2 Klein 4B	4B	~14 GB	~22 GB
FLUX.2 Klein 9B	9B	~17 GB	~32 GB
FLUX.1 Schnell	12B	~18 GB	~38 GB
FLUX.1 Dev	12B	~18 GB	~38 GB
FLUX.2 Dev	32B	~28 GB	~78 GB

NF4 is enabled by default (use_nf4: true). Models exceeding available VRAM will use CPU offload automatically. Only Klein 4B fits a 16 GB GPU without offloading.

Troubleshooting

Problem	Solution
CUDA out of memory	Use a smaller model (klein-4b), reduce resolution, or enable NF4 in config
401 / gated / token error	Set HF_TOKEN in .env and accept the model license on HuggingFace
Module not found	Run pip install -r requirements.txt
Slow generation	Enable NF4 (use_nf4: true), reduce steps, or use klein-4b
Model not downloading	Check internet connection or if original model links have been removed

License

This tool itself is provided as-is under the MIT License. Each model has its own license:

• Apache 2.0: klein-4b, schnell — free for commercial use
• FLUX Non-Commercial: dev, klein-9b, flux2-dev — research/personal use only

You are responsible for complying with each model's license terms.