Sd35l

sd35l - Stable Diffusion 3.5 Large Image Generation

Overview

Generate high-quality images using Stability AI SD3.5 Large on Amazon Bedrock. SD3.5 Large uses the MMDiT (Multi-Modal Diffusion Transformer) architecture, delivering strong prompt adherence, photorealistic quality, and improved text rendering over previous Stable Diffusion models.

Model: stability.sd3-5-large-v1:0 Region: us-west-2 (direct invocation, no cross-region inference) API: InvokeModel (bedrock-runtime) with Stability AI format Status: GA (Generally Available)

Workflow

Step 1: Analyze User Request

Determine the task type from the user's request:

| User Intent | Mode | |-------------|------| | Generate new image from text | text-to-image | | Transform existing image style / create variation | image-to-image |

Step 2: Optimize the Prompt

Transform the user's raw input into an optimized prompt. Consult references/prompt_best_practices.md for the full transformation rules.

Core transformation rules:

1. Caption style - Convert commands ("make X", "draw X") to descriptive captions
2. Six-element structure - Subject, Environment, Pose/Action, Lighting, Camera, Style
3. Negative prompt separation - Extract negation words into negative_prompt
4. Korean to English - Translate Korean input to English, preserving visual intent
5. Concrete language - Replace vague terms with specific visual descriptors

Step 3: Determine Parameters

| Parameter | Default | Notes | |-----------|---------|-------| | aspectratio | 1:1 | Options: 1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9, 9:21 | | seed | random | 0-4294967294 for reproducibility | | outputformat | png | png or jpeg | | mode | text-to-image | text-to-image or image-to-image | | strength | 0.7 | image-to-image only, 0.0-1.0 (higher = more change) |

Step 4: Present Plan for Confirmation

Before calling the API, show the user:

## Image Generation Plan

**Model**: SD3.5 Large
**Optimized Prompt**: [transformed prompt in English]
**Negative Prompt**: [extracted negative elements, if any]
**Aspect Ratio**: 1:1
**Seed**: [random or specified]

Proceed with generation?

Wait for user confirmation before proceeding.

Step 5: Generate Image

Run the generation script:

python3 ${SKILL_PATH}/scripts/generate_image.py \
  --prompt "optimized prompt here" \
  --negative-prompt "negative elements here" \
  --aspect-ratio 1:1 \
  --seed 42 \
  --output-dir ./output

For image-to-image:

python3 ${SKILL_PATH}/scripts/generate_image.py \
  --prompt "description of desired result" \
  --image path/to/source.png \
  --strength 0.7 \
  --output-dir ./output

Step 6: Present Results

After generation, display the result:

• Show the saved file path
• Show the seed used (for reproducibility)
• Read and display the generated image using the Read tool
• Offer to regenerate with a different seed or adjusted prompt

Task-Specific Instructions

Text-to-Image

• Apply the full prompt optimization pipeline
• Default to 1:1 aspect ratio unless user specifies otherwise
• SD3.5 excels with detailed, descriptive prompts — front-load important elements

Image-to-Image (Style Transfer / Variation)

• Require an existing image path from the user
• strength controls how much the output differs from the input:
• 0.1-0.3: Subtle changes, preserves most of the original
• 0.4-0.6: Moderate transformation
• 0.7-0.9: Major changes, only composition preserved
• 1.0: Completely new image guided only by the prompt
• Describe the desired result in the prompt, not the changes

Negative Prompts

• Extract negation words from user input into negative_prompt
• Common quality negatives: "blurry, distorted, low quality, bad anatomy, disfigured"
• SD3.5 handles negative prompts well — use them to refine output quality

Additional Resources

Reference Files

• references/prompt_best_practices.md - Complete prompt optimization rules, examples, templates
• references/api_reference.md - SD3.5 Large API parameters and response format

Scripts

• scripts/generate_image.py - Main image generation script supporting text-to-image and image-to-image

Error Handling

| Error | Cause | Solution | |-------|-------|----------| | Content filter | Prompt triggers safety filter | Revise prompt; check finish_reasons for "CONTENT_FILTERED" | | Access denied | Model not enabled | Enable SD3.5 Large in Bedrock console for us-west-2 | | Timeout | Generation taking too long | Script sets 120s timeout | | Invalid aspect ratio | Unsupported ratio | Use one of: 1:1, 16:9, 9:16, 4:3, 3:4, 2:3, 3:2, 21:9, 9:21 |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jesamkim
• Source: jesamkim/oh-my-skills
• License: MIT

Install and usage instructions live in the source repository linked above.