Moark Image Gen

Generate high-quality images from text descriptions.

Image Generator

This skill allows users to generate high-quality images based on text descriptions using an external image generation API(Gitee AI).

Usage

Ensure you have installed the required dependencies (pip install openai). Use the bundled script to generate images.

Qwen-Image (Default)

bash

python {baseDir}/scripts/perform_image_gen.py --prompt "your image description" --model Qwen-Image --size 1024x1024 --negative-prompt "elements to avoid" --num-inference-steps 30 --api-key YOUR_API_KEY

Kolors

bash

python {baseDir}/scripts/perform_image_gen.py --prompt "your image description" --model Kolors --size 1024x1024 --num-inference-steps 25 --guidance-scale 7.5 --api-key YOUR_API_KEY

GLM-Image

bash

python {baseDir}/scripts/perform_image_gen.py --prompt "your image description" --model GLM-Image --size 1024x1024 --negative-prompt "elements to avoid" --num-inference-steps 30 --guidance-scale 1.5 --api-key YOUR_API_KEY

HunyuanDiT-v1.2-Diffusers-Distilled

bash

python {baseDir}/scripts/perform_image_gen.py --prompt "your image description" --model HunyuanDiT-v1.2-Diffusers-Distilled --size 1024x1024 --negative-prompt "elements to avoid" --num-inference-steps 25 --guidance-scale 5.0 --api-key YOUR_API_KEY

FLUX.2-dev

bash

python {baseDir}/scripts/perform_image_gen.py --prompt "your image description" --model FLUX.2-dev --size 1024x1024 --negative-prompt "elements to avoid" --num-inference-steps 20 --guidance-scale 7.5 --api-key YOUR_API_KEY

Options

Sizes:

• 256x256 - Small square format
• 512x512 - Square format
• 1024x1024(default) - Square format
• 1024x576 - 16:9 landscape
• 576x1024 - 9:16 portrait
• 1024x768 - 4:3 format
• 768x1024 - 3:4 portrait
• 1024x640 - 16:10 landscape
• 640x1024 - 10:16 portrait
• 2048x2048 - High-resolution square format

Additional flags:

• --model - Specify the model to use. Options include Qwen-Image (default), Kolors, GLM-Image, FLUX.2-dev, HunyuanDiT-v1.2-Diffusers-Distilled.
• --negative-prompt - Specify what elements users want to avoid in the generated image(default: "低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，人脸无细节，过度光滑，画面具有AI感。构图混乱。文字模糊，扭曲。").
• --size - Specify the size of the generated image. Options include 256x256, 512x512, 1024x1024 (default), 1024x576, 576x1024, 1024x768, 768x1024, 1024x640, 640x1024, 2048x2048.
• --guidance-scale - Float value to control how closely the model adheres to the prompt (default depends on model).
• --num-inference-steps - Integer for denoise steps (default depends on model). Higher values typically increase quality but take longer.

Model Specific Defaults:

• Kolors: steps 25 (range 20-30), scale 7.5 (range 0-100)
• Qwen-Image: steps 30 (range 4-50)
• GLM-Image: steps 30 (range 10-50), scale 1.5 (range 0-10)
• HunyuanDiT-v1.2-Diffusers-Distilled: steps 25 (range 25-50), scale 5 (range 0-20)
• FLUX.2-dev: steps 20 (range 10-50), scale 7.5 (range 0-100)

Workflow

1. Execute the perform_image_gen.py script with the parameters from the user.
2. Parse the script output and find the line starting with IMAGE_URL:.
3. Extract the image URL from that line (format: IMAGE_URL: https://...).
4. Display the image to the user using markdown syntax: 🖼️[Generated Image](URL).

Notes

• You should not only return the image URL but also describe the image based on the user's prompt, and claim the hyperparameters used for generation.
• You should always wait for the script to finish executing, don't shut it down prematurely.
• The Lanaguage of your answer should be consistent with the user's question.
• By default, return image URL directly without downloading.
• If GITEEAI_API_KEY is none, the user must provide --api-key argument.
• The script prints IMAGE_URL: in the output - extract this URL and display it using markdown image syntax: 🖼️[Generated image](URL).
• Always look for the line starting with IMAGE_URL: in the script output and render the image for the user.
• You should honestly repeat the description of the image from user without any additional imaginations.
• Handling User Feedback on Quality: If the user states the image quality is low or lacks details, you should retry generating with a higher --num-inference-steps (e.g. 25 → 30).
• Handling User Feedback on Prompt Adherence: If the user states the image doesn't follow the prompt closely enough or ignores details, increase the --guidance-scale parameter (e.g. 7.5 → 15). If they say it's oversaturated or distorted, decrease it.