Sd prompt forge

SD Prompt Forge

Build prompts as reproducible generation specifications instead of unstructured tag lists.

Workflow

1. Identify the target model, interface, subject count, aspect ratio, and available adapters.
2. Separate global scene requirements from character-specific requirements.
3. Build the positive prompt in the order below.
4. Add a focused negative prompt based on likely failure modes.
5. Recommend weights and generation settings as starting points, not guarantees.
6. For multiple characters, define regions and consistency controls explicitly.
7. Return the prompt, negative prompt, settings, and a short rationale.

Positive Prompt Order

Use this order unless the target model documents a different convention:

[model or style adapters]
[character LoRAs and activation tags]
[quality and medium]
[subject count and composition]
[environment and camera]
[character identity and appearance]
[clothing and accessories]
[pose, expression, and interaction]
[lighting, atmosphere, and final details]

Keep related tags together. Prefer concrete visual attributes over abstract adjectives.

LoRA Guidance

• Place each character LoRA near that character's identity tags.
• Start character LoRAs around 0.65-0.85.
• Start style LoRAs around 0.30-0.60.
• Reduce competing LoRAs before increasing prompt weights.
• Include activation tags immediately after their LoRA.
• Explain any weight above 1.0; high weights often damage anatomy or style balance.

Do not assume a LoRA name or activation tag. Ask for it or use a clear placeholder.

Negative Prompts

Build negatives from the requested image, not from one universal block.

Always consider:

• anatomy and hand failures
• duplicate subjects or limbs
• text, signatures, and watermarks
• cropping or framing mistakes
• unwanted style or medium
• attributes that could bleed between characters

Avoid long contradictory negatives. If the user requests an attribute, do not also negate it.

Multi-Character Scenes

Keep global tags before regional character blocks:

[global scene, quality, camera, background]
AND
[character A, left-side position, pose, identity]
AND
[character B, right-side position, pose, identity]

For detailed regional layouts, color-bleed prevention, and consistency adapters, read references/multi-character-scenes.md.

Control Tools

Choose the smallest useful control stack:

• OpenPose for body layout and interaction.
• IP-Adapter for identity or appearance consistency.
• Depth for spatial structure.
• Lineart or Canny for silhouette and edge control.
• ADetailer for face or hand correction after generation.

Avoid stacking controls that solve the same problem unless the user has tested the combination.

Model Adaptation

• SDXL / Illustrious XL: use descriptive natural language mixed with recognized tags.
• Pony models: follow the checkpoint's score and source-tag conventions when documented.
• Danbooru-trained models: prefer known tag vocabulary and concise attribute groups.
• Unknown checkpoints: provide a conservative prompt and label model-specific assumptions.

Read references/prompt-patterns.md for reusable solo, product-shot, and scene templates.

Output Format

Return:

## [Image or scene name]

**Positive prompt**
[prompt]

**Negative prompt**
[negative prompt]

**Suggested settings**
- Model:
- Sampler:
- Steps:
- CFG:
- Resolution:
- Adapters:

**Why this structure**
[2-4 concise bullets]

When the user provides test results, revise one variable group at a time: composition, identity, style, anatomy, then detail.

Responsible Use

Respect model licenses, artist and character rights, platform rules, and consent requirements. Do not present generated images as authentic photographs or evidence.