Prompt Engineering

ComfyUI Prompt Engineering

CLIP Text Encoding Fundamentals

ComfyUI uses CLIP (Contrastive Language-Image Pre-training) text encoders to convert text prompts into conditioning tensors. The CLIPTextEncode node takes a text string and a CLIP model, producing a CONDITIONING output for the KSampler.

Token Limit

CLIP processes text in 77-token chunks. Each word is typically 1-3 tokens. Prompts exceeding 77 tokens are silently truncated unless you use the BREAK token or a multi-clip encoding node.

Weight Syntax

Emphasis (Attention Weights)

Adjust how strongly the model attends to specific words or phrases:

Weight Rules

• Valid range: 0.0 to 2.0 (going beyond 1.5 often causes artifacts)
• Default weight: 1.0 for unmodified tokens
• Nesting stacks multiplicatively: ((word)) = 1.1 * 1.1 = (word:1.21)
• Phrases: (red sports car:1.3) applies weight to the entire phrase
• Mixing: (detailed face:1.4), (blurry background:0.6), combined in one prompt

Examples

a (beautiful:1.3) woman with (flowing red hair:1.2), wearing a blue dress, (sharp focus:1.1)

(masterpiece:1.4), (best quality:1.3), a knight in (ornate armor:1.2), standing on a cliff, (dramatic lighting:1.1), cinematic

BREAK Token

The BREAK keyword forces CLIP to end the current 77-token chunk and start processing subsequent text in a new chunk. This is critical for long prompts.

When to Use BREAK

• Prompt exceeds ~60 words (approaching the 77-token limit)
• You want to separate conceptually distinct parts of the prompt
• Certain details are being ignored (they may be past the 77-token cutoff)

BREAK Example

masterpiece, best quality, a beautiful Japanese garden with cherry blossoms,
stone lanterns, koi pond, traditional wooden bridge, morning mist
BREAK
highly detailed, 8k uhd, photorealistic, volumetric lighting,
depth of field, golden hour, award-winning photography

Each chunk is encoded independently and then concatenated as conditioning, so all tokens are processed.

Embeddings / Textual Inversions

Embeddings (textual inversions) are pre-trained token sets that encode complex concepts into a single trigger word.

Syntax

embedding:easynegative
embedding:badhandv4
embedding:bad-image-v2-39000

Usage in Prompts

• Place embedding triggers directly in the prompt text
• Most commonly used in negative prompts to improve quality
• The embedding .safetensors or .pt file must be in models/embeddings/

Common Negative Embeddings

Example with Embeddings

Positive: a portrait of a woman, masterpiece, best quality Negative: embedding:easynegative, embedding:badhandv4, worst quality, low quality

Model-Specific Prompting

SD 1.5

Negative prompt: IMPORTANT. SD 1.5 is sensitive to negatives.

Positive prompt structure:

(masterpiece:1.2), (best quality:1.2), subject description, details, style tags

Recommended negative prompt:

worst quality, low quality, normal quality, lowres, watermark, signature,
text, jpeg artifacts, blurry, bad anatomy, bad hands, extra fingers,
missing fingers, extra limbs, deformed, disfigured, mutation, ugly

Key notes:

• Quality tags like masterpiece, best quality make a large difference to output
• Responds well to danbooru-style tags: 1girl, long hair, blue eyes, school uniform
• Embedding-based negatives (easynegative) work well
• Keep prompts concise, since each chunk is limited to 77 tokens

SDXL (1.0 / Turbo / Lightning)

Negative prompt: Moderate importance. SDXL is less sensitive to negatives than SD 1.5.

Positive prompt structure:

subject description with natural language, detailed description of scene and style

Recommended negative prompt:

blurry, low quality, deformed, ugly, bad anatomy, disfigured, poorly drawn face,
mutation, mutated, extra limbs, watermark, text

Key notes:

• SDXL understands natural language better than tag-based prompts
• Dual CLIP encoders (CLIP-L + CLIP-G). Use CLIPTextEncodeSDXL for separate control
• CLIPTextEncodeSDXL has separate text_g (global description) and text_l (local details) fields
• Supports longer prompts natively (two 77-token chunks via dual CLIP)
• Quality tags are less critical but still helpful
• SDXL Turbo: 1-4 steps, CFG 1.0-2.0, minimal negative prompt needed
• SDXL Lightning: 4-8 steps, CFG 1.0-2.0, often works with empty negative

Flux (Flux.1 schnell / dev)

Negative prompt: NOT USED. Flux operates at CFG=1.0 with no negative conditioning.

Positive prompt structure:

Detailed natural language description. Flux excels with descriptive sentences
rather than comma-separated tags. Describe the scene as if writing a paragraph.

Key notes:

• CFG must be 1.0. Higher values cause artifacts
• No negative prompt. Connect nothing or empty string to negative conditioning
• T5-XXL encoder understands complex sentences and spatial relationships
• Flux handles compositional prompts better than SD models
• Longer prompts (200+ tokens) work well thanks to T5 encoder
• Prompt structure: describe the scene naturally, like a caption
• Schnell: 4 steps, simple scheduler
• Dev: 20-50 steps, sgm_uniform scheduler

Flux Prompt Example

A serene Japanese garden in autumn. A stone path leads through a grove of maple
trees with bright red and orange leaves. A small wooden bridge crosses a koi pond
where golden fish swim beneath the surface. Morning mist rises from the water,
and soft sunlight filters through the canopy. The scene is photorealistic with
warm, natural lighting and shallow depth of field.

SD3 / SD3.5

Negative prompt: Minimal. SD3 needs little negative guidance.

Positive prompt structure:

Natural language description, supports very long detailed prompts thanks to T5-XXL

Key notes:

• Triple CLIP architecture: CLIP-L + CLIP-G + T5-XXL
• Supports much longer prompts than SD 1.5 or SDXL
• Natural language works better than tag-based prompting
• CFG 4-7 (lower than SD 1.5)
• Minimal negatives needed. low quality, blurry is usually sufficient
• Use CLIPTextEncodeSD3 node for model-specific encoding if available

Prompt Structure Best Practices