image-analysis-sd 0.1.0

Reverse Stable Diffusion utilities and CLI

Image Description Tool

Repository: https://github.com/chazzofalf/image_analysis_and_sd

This tool leverages Ollama models to describe images and generate stable diffusion parameters for recreation in InvokeAI.

Functionality

• Image Description: Uses multiple Ollama models (gemma3:27b, llava:latest, llama4:scout, mistral-small3.2:24b, Moondream:latest, llama3.2-vision:latest) to generate descriptions of an input image.
• Summarization: Combines the individual descriptions into a single, comprehensive summary using gpt-oss:120b.
• Stable Diffusion Parameter Generation: Extracts parameters (positive prompt, negative prompt, step count, CFG scale, image dimensions, and suggested model/scheduler) suitable for recreating the image in InvokeAI, using gpt-oss:120b.

Usage

1. Ensure you have Ollama installed and configured.

2. Run the image_describer.py script with the following arguments:

   python image_describer.py <image_path> <output_file>
   

   • <image_path>: Path to the input image.
   • <output_file>: Path to the output file (summary will be written to this file, and stable diffusion parameters will be written to <output_file>.sd.txt).

Installation

1. Clone the repository:

   git clone https://github.com/chazzofalf/image_analysis_and_sd
   cd image_analysis_and_sd
   

2. Create a virtual environment:

   python3 -m venv venv
   

3. Activate the virtual environment:
   • Linux/macOS:

     source venv/bin/activate
     

   • Windows:

     .\venv\Scripts\activate
     

4. Install dependencies:

   pip install -r requirements.txt
   

Example

Here's an example using the Scout.jpg image:

• Image: Scout.jpg
• Description: Scout.jpg.txt
• Stable Diffusion Instructions: Scout.jpg.txt.sd.txt

Quick-Start “InvokeAI” Settings

Item	Recommendation
Positive Prompt	a small dark-fur animal curled up sleeping on a cozy patterned blanket, realistic photography, soft natural light filtered through a sheer white curtain, navy-blue blanket with tiny white/ light-blue four-petal flower motif, lightly rumpled sky-blue sheet underneath, plain white wall background, subtle window with diffused daylight, shallow depth of field, ultra-sharp focus on animal’s face, warm-amber glow, highly detailed fur, plush texture, intimate bedroom scene, 8k resolution, cinematic composition, top-down angle
Negative Prompt	low-resolution, blurry, jpeg artifacts, over-exposed, under-exposed, color banding, chromatic aberration, watermark, text, signature, frame, lens flare, cartoon, illustration, anime, painting, oil-painting, stylized, vignette, grainy, noise, over-sharpened, HDR, unrealistic shadows, missing limbs, extra paws, duplicated body parts, deformed anatomy, unrealistic proportions, plastic look
Steps	35–45 (35 works well for most cases; push to 45 if you need extra fine detail on fur and fabric)
CFG Scale	7.5–9 (7.5 for a balanced blend of prompt fidelity and creativity; 9 if you want the image to follow the prompt almost verbatim)
Suggested Image Dimensions	768 × 1024 (portrait – good for a bedside view) or 1024 × 1280 if you want a bit more vertical space for the window and wall.
Best Stable-Diffusion Model	Stable Diffusion XL 1.0 (or XL 1.0‑refiner) – it handles realistic lighting, fine fur texture and fabric patterns far better than the 1.5 checkpoint. If you have limited VRAM, the SDXL‑Turbo checkpoint can also produce good results at a lower cost, but expect slightly less micro‑detail.
Best Scheduler	DPM++ 2M Karras (very stable for photorealistic output) – alternatively Euler‑a works well for a smoother, less “noisy” finish.

How the Prompt Was Built (Why It Works)

• Subject Detail – “small dark‑fur animal” covers both the dog‑or‑cat ambiguity while keeping the description concise for the model. Adding “curled up sleeping” forces the pose.
• Fabric & Pattern – Explicitly naming the navy‑blue blanket and tiny white/light‑blue four‑petal flower motif gives the model a concrete visual cue for the patterned textile, which is often a weak point for generic prompts.
• Lighting & Atmosphere – Phrases like “soft natural light filtered through a sheer white curtain”, “warm‑amber glow”, and “diffused daylight” steer the model toward the low‑key, gently illuminated mood described.
• Depth & Focus – “shallow depth of field”, “ultra‑sharp focus on animal’s face”, and “top‑down angle” ensure the animal is the clear focal point while the background stays softly rendered.
• Quality Tags – “realistic photography”, “8k resolution”, “cinematic composition” push the model toward a high‑detail, photorealistic output rather than an illustration or stylized rendering.

Quick “Copy‑Paste” Block for InvokeAI

prompt: |
  a small dark-fur animal curled up sleeping on a cozy patterned blanket, realistic photography, soft natural light filtered through a sheer white curtain, navy-blue blanket with tiny white light-blue four-petal flower motif, lightly rumpled sky-blue sheet underneath, plain white wall background, subtle window with diffused daylight, shallow depth of field, ultra-sharp focus on animal’s face, warm-amber glow, highly detailed fur, plush texture, intimate bedroom scene, 8k resolution, cinematic composition, top-down angle
negative_prompt: |
  low-resolution, blurry, jpeg artifacts, over-exposed, under-exposed, color banding, chromatic aberration, watermark, text, signature, frame, lens flare, cartoon, illustration, anime, painting, oil-painting, stylized, vignette, grainy, noise, over-sharpened, HDR, unrealistic shadows, missing limbs, extra paws, duplicated body parts, deformed anatomy, unrealistic proportions, plastic look
steps: 40
cfg_scale: 8
width: 768
height: 1024
sampler: DPM++ 2M Karras
model: stabilityai/stable-diffusion-xl-base-1.0   # or the XL‑Turbo checkpoint if VRAM is limited

Feel free to tweak the CFG (higher for stricter adherence) or steps (more steps for extra fine fur detail). Adjust the width/height if you prefer a wider composition that includes more of the window or wall.

Happy generating! 🎨🖼️

Dependencies

• Ollama
• Python 3.10.18

License

This project is licensed under the MIT License - see the LICENSE file for details.

Made with Cline and Ollama, using models gpt-oss:120b and gemma3:27b.