comfyui-prompt-engineer（comfyui 提示工程师）

ComfyUI Prompt Engineer

Generates optimized prompts tailored to specific models and identity methods. Different models respond differently to prompts.

Model-Specific Prompt Rules

FLUX.1 (dev/schnell/Kontext)

• Style: Natural language descriptions work best
• CFG: 3.5-4 (very low)
• Quality tags: Minimal - FLUX doesn't need "masterpiece, best quality"
• Length: Medium (50-100 words)
• Structure: {subject description}, {setting}, {lighting}, {camera/style}

Good FLUX prompt:

photorealistic portrait of a woman with auburn hair and green eyes, freckles across
her nose and cheeks, wearing a cream knit sweater, sitting in a cafe with warm ambient
lighting, shallow depth of field, shot on Sony A7IV, 85mm lens

Bad FLUX prompt (too many quality tags):

masterpiece, best quality, 8k uhd, highly detailed, photorealistic portrait...

SDXL (RealVisXL, Juggernaut, etc.)

• Style: Quality tags at front help significantly
• CFG: 7-9
• Quality tags: Include masterpiece, best quality, photorealistic
• Length: Medium-long (50-150 words)
• Structure: {quality tags}, {trigger word}, {subject}, {details}, {setting}, {style}
• Weighted syntax: Supported (important:1.3) or ((very important))

Good SDXL prompt:

masterpiece, best quality, sage_character, photorealistic portrait of a woman,
detailed skin texture with freckles, emerald green eyes, auburn copper hair,
natural lighting from window, indoor setting, shallow depth of field,
RAW photo quality, 8k uhd, film grain

SD 1.5

• Style: Tag-based works best
• CFG: 7-8
• Quality tags: Essential
• Length: Shorter (30-80 words)
• Structure: {quality}, {trigger}, {subject}, {details}, {style tags}

Wan 2.1/2.2 (Video)

• Style: Concise motion descriptions
• CFG: 5-7
• Quality tags: Minimal
• Length: Short (20-50 words)
• Focus: Describe the motion, not just the appearance
• Structure: {subject}, {action/motion}, {setting}, {quality}

Good Wan prompt:

young woman with auburn hair, talking naturally with gentle hand gestures,
seated at a modern desk, soft studio lighting, high quality

AnimateDiff

• Style: Same as base model (SD1.5/SDXL) but add motion keywords
• Motion keywords: Include camera/subject motion description
• Length: Same as base model

With Identity Methods

InstantID Prompts

Key adjustments:

• DO NOT describe specific facial features (the model provides them)
• DO describe everything else: clothing, pose, setting, lighting
• Keep CFG at 4-5

photorealistic portrait, wearing black leather jacket, standing in an alley,
dramatic side lighting, urban setting, moody atmosphere, 8k quality

PuLID Prompts

Key adjustments:

• Can include some facial descriptions (PuLID is more tolerant)
• Method "neutral" for realistic, "fidelity" for exact match
• CFG 5-7

sage_character, photorealistic portrait, green eyes visible, natural makeup,
professional headshot, neutral grey background, studio lighting

IP-Adapter FaceID

Key adjustments:

• Describe the style you want, not the face
• weight_type "style transfer" for 3D→realistic
• Standard CFG for base model

photorealistic portrait, DSLR photo quality, natural skin texture,
warm indoor lighting, bokeh background, professional photography

FLUX Kontext (Editing)

Key adjustments:

• Describe the EDIT, not the full image
• Be specific about what to change
• Mention what to preserve

Change the outfit to a formal black evening dress while keeping the face,
hair, and pose exactly the same. Add subtle jewelry.

With Character LoRA

Key adjustments:

• ALWAYS include trigger word first
• Don't describe features the LoRA has learned
• Focus on what varies: pose, clothing, setting, lighting

sage_character, standing on a rooftop at sunset, wind blowing hair,
wearing casual summer dress, city skyline behind, golden hour lighting,
cinematic composition

Negative Prompt Templates

Universal Negative (SDXL/SD1.5)

(worst quality:1.4), (low quality:1.4), blurry, deformed, bad anatomy,
bad hands, extra fingers, missing fingers, extra limbs, fused fingers,
text, watermark, signature, jpeg artifacts, username, error

Photorealism Negative

3d render, cartoon, anime, illustration, painting, drawing, cgi,
plastic skin, smooth skin, airbrushed, video game, doll, mannequin,
oversaturated, artificial lighting

Video Negative (Wan/AnimateDiff)

static, frozen, jerky motion, low quality, blurry, distorted face,
bad anatomy, glitch, artifacts, flickering, jittery, unnatural movement

FLUX Negative (keep minimal)

blurry, low quality, distorted, deformed, ugly, watermark, text

Prompt Construction Workflow

1. Get character profile from projects/{project}/characters/{name}/profile.yaml
2. Get target model from inventory/user preference
3. Apply model rules from above
4. Add identity method modifiers if applicable
5. Include trigger word if LoRA is being used
6. Draft positive + negative pair
7. Review against past successes in character's generation_history
8. Adjust CFG recommendation based on method stack

Reference

• references/prompt-templates.md - Full template library with examples
• references/workflows.md - CFG and sampler settings per workflow
• Character profiles in projects/ for trigger words and feature descriptions