ImgGenHub

A personal, cost-efficient AI image generation platform.

ImgGenHub is a personal image generation hub that connects to web-based image generation services, currently featuring a fully automated Kaggle-based pipeline with plans for multi-platform support.

---

Features

Current functionality

Kaggle-powered image generation

• Automated pipeline: Deploy → Monitor → Download workflow
• Automated secret management: Local .env secret detected and auto-uploaded to Kaggle dataset
• Multiple models: Supports Stable Diffusion variants, Flux.1-schnell GGUF quantized (Q4) version and Flux.1-schnell bf16 version.

Vast.ai support (todo)

Planned in Dec 2025 / Jan 2026.

---

Requirements

• 3.11+ Python and Poetry
• Kaggle Account: with API credentials
• HuggingFace token: for accessing gated models (e.g., FLUX.1-schnell)

---

Usage

Using FLUX.1-schnell

FLUX.1-schnell is commercially usable. The full fp32 model doesn't fit into our available VRAM on Kaggle kernels but its lossless bf16 compression does.

FLUX.1-schnell with bf16

poetry run imggenhub \
  --prompt "photorealistic indoor restaurant scene" \
  --model_id "black-forest-labs/FLUX.1-schnell" \
  --steps 4 \
  --guidance 0.75 \
  --img_width 1024 \
  --img_height 1024 \
  --precision bf16 \
  --gpu

Multiple prompts via command line

poetry run imggenhub \
  --prompt "A serene mountain landscape at sunset" \
  --prompt "A bustling city street in the rain" \
  --prompt "An abstract geometric pattern" \
  --model_id "black-forest-labs/FLUX.1-schnell" \
  --steps 4 \
  --guidance 0.75 \
  --img_width 1024 \
  --img_height 1024 \
  --precision bf16 \
  --gpu

FLUX.1-schnell quantized version (Q4_0 GGUF)

poetry run imggenhub \
  --prompt "photorealistic indoor restaurant scene" \
  --model_id "city96/FLUX.1-schnell-gguf" \
  --model_filename "flux1-schnell-Q4_0.gguf" \
  --steps 4 \
  --guidance 0.8 \
  --img_width 512 \
  --img_height 512 \
  --precision q4 \
  --gpu

Stable diffusion XL with refiner

poetry run imggenhub \
  --prompt "photorealistic indoor restaurant scene" \
  --model_id stabilityai/stable-diffusion-xl-base-1.0 \
  --refiner_model_id stabilityai/stable-diffusion-xl-refiner-1.0 \
  --steps 30 \
  --guidance 8.0 \
  --precision fp16 \
  --refiner_steps 15 \
  --refiner_guidance 7.0 \
  --refiner_precision fp16 \
  --gpu

Supported flags

General flags (all models)

• --prompt: Single prompt or multiple prompts (use flag multiple times)
• --prompts_file: JSON file with multiple prompts
• --gpu: Enable GPU acceleration (required for FLUX.1 models)
• --steps: Inference steps (50-100 for Stable Diffusion models, ~4 for FLUX)
• --guidance: Prompt adherence strength (7-12 recommended for photorealism, 0.75-1.0 for FLUX)
• --precision: Model precision (fp32/fp16/int8/int4 for base models; q4/q5/q6 for GGUF quantized models)
• --img_width / --img_height: Image dimensions (must be multiples of 64)
• --negative_prompt: Quality control prompts
• --dest DEST: Custom name prefix for the output folder (default: timestamp only)
• --notebook: Custom notebook path
• --kernel_path: Kaggle kernel configuration directory
• Outputs are always saved under output/ with automatic timestamping
• All logs from the generation process are saved in the same folder as the images

Stable Diffusion XL flags (model_id: "stabilityai/*")

• --model_id: Hugging Face model ID (e.g., stabilityai/stable-diffusion-xl-base-1.0)
• --refiner_model_id: SDXL refiner model for enhanced photorealism (e.g., stabilityai/stable-diffusion-xl-refiner-1.0)
• --refiner_steps: Inference steps for refiner (defaults to 20)
• --refiner_guidance: Guidance scale for refiner (defaults to same as --guidance)
• --refiner_precision: Precision for refiner (defaults to same as --precision)
• --refiner_negative_prompt: Negative prompt for refiner (defaults to same as --negative_prompt)

FLUX.1-schnell bf16 flags (model_id: "black-forest-labs/FLUX.1-schnell")

• --model_id: Must be "black-forest-labs/FLUX.1-schnell"

Note: Refiner-related flags are ignored for FLUX models

FLUX.1-schnell GGUF (quantized) flags (model_id: "city96/FLUX.1-schnell-gguf")

• --model_id: Hugging Face repository containing quantized model (e.g., city96/FLUX.1-schnell-gguf)
• --model_filename: GGUF model filename (e.g., flux1-schnell-Q4_0.gguf)
• --vae_repo_id: VAE model repository (auto-resolved if not specified)
• --vae_filename: VAE model filename (auto-resolved if not specified)
• --clip_l_repo_id: CLIP-L component repository (auto-resolved if not specified)
• --clip_l_filename: CLIP-L component filename (auto-resolved if not specified)
• --t5xxl_repo_id: T5-XXL component repository (auto-resolved if not specified)
• --t5xxl_filename: T5-XXL component filename (auto-resolved if not specified)

Note: Refiner-related flags are ignored for FLUX models

Custom Kaggle datasets

To speed up inference, we upload the largest model files as a custom Kaggle dataset. This massively speeds up model inference as there is no need to fetch tens of GBs before every inference run.

Note: temporarily switched off as we found that for FLUX.1-schnell models (and hence, likely for any large model) it actually takes more time to read these models from Kaggle datasets inside of a Kaggle notebook than downloading directly from Hugging Face.

Requesting specific GPU types on Kaggle

Kaggle kernels accept an optional accelerator field in kernel-metadata.json. This repository's Kaggle connector provides a convenient way to set that field when creating a deployment. Two accelerator values are supported in our tooling:

• nvidia-t4-x2 — request two NVIDIA T4 GPUs (T4 x2)
• nvidia-p100 — request a single NVIDIA P100 GPU (P100)

Important notes:

• Kaggle may not always honor a specific accelerator if capacity is limited; it can fall back to an available GPU type.
• The accelerator field is set through the deployment metadata and pushed with the kernel.

Examples — using the provided example deployer (located at kaggle_gpu_connector/examples/deploy_custom.py):

Request a P100 GPU:

poetry run python kaggle_gpu_connector/examples/deploy_custom.py nvidia-p100

Request T4 x2 GPUs:

poetry run python kaggle_gpu_connector/examples/deploy_custom.py nvidia-t4-x2

What the example does:

• Creates a temporary kernel-metadata.json with the chosen accelerator value.
• Converts the example Python script (kaggle_gpu_connector/examples/custom_job.py) to a notebook and uploads it as a Kaggle kernel.
• Waits for resources, pushes the kernel, polls until completion, and then you can download the kernel output with the Kaggle CLI:

poetry run kaggle kernels output <your-username>/<kernel-id> -p outputs/custom_test

If you want to manually create or edit metadata, ensure your kernel-metadata.json includes:

{
  "id": "username/your-kernel",
  "code_file": "your_notebook.ipynb",
  "kernel_type": "notebook",
  "enable_gpu": true,
  "accelerator": "nvidia-t4-x2"
}

See kaggle_gpu_connector/examples/deploy_custom.py for a minimal programmatic example of creating metadata and deploying with either accelerator.