Overview
Generates optimized prompts tailored to a specific model and identity method. Diffusion checkpoints do not share a prompt language. Each family was trained with a different text encoder on a different caption distribution, so the same sentence lands differently depending on the target. This skill picks the rules that match the checkpoint and the identity-preservation method actually in the stack.
When to Use
Activate this skill when the job-to-be-done is crafting or refining the prompt text for a ComfyUI generation. Triggers on requests to:
- Generate a new positive/negative prompt pair for a specific checkpoint (FLUX, SDXL, SD1.5, Wan, AnimateDiff)
- Refine or debug an existing prompt that is underperforming on a given model
- Adapt a prompt when switching checkpoints (e.g., an SDXL prompt that needs to run on FLUX)
- Tailor a prompt to an identity method (InstantID, PuLID, IP-Adapter FaceID, FLUX Kontext, character LoRA)
- Build a model-appropriate negative prompt
- Recommend a CFG range that matches the model + identity-method stack
- Include the correct trigger word and feature handling when a character LoRA is in play
Prerequisites
- ComfyUI installed and target checkpoint identified.
- Character profile available at
projects/{project}/characters/{name}/profile.yaml (if using a character LoRA or specific identity).
- Optional: the Examples section below for fuller prompt pairs; the model-rule tables in Procedure for CFG ranges. This skill crafts prompt text for a known checkpoint — it does not interview a vague idea (use
comfyui-prompt-interview for that).
Procedure
1. Identify Target Model and Identity Method
Determine the checkpoint family (FLUX, SDXL, SD1.5, Wan, AnimateDiff) and the identity method (InstantID, PuLID, IP-Adapter FaceID, FLUX Kontext, Character LoRA) in use.
2. Load Character Profile
If using a character LoRA or specific identity, read the character profile from projects/{project}/characters/{name}/profile.yaml to get the trigger word and canonical feature list. Do not invent details.
3. Apply Model-Specific Rules
FLUX.1 (dev / schnell / Kontext)
- Style: Natural-language descriptions. FLUX uses a T5-XXL text encoder alongside CLIP and parses sentences rather than comma-separated booru tags.
- CFG: 3.5–4 (very low) for
dev; schnell runs near CFG 1. High CFG over-saturates color and "fries" the image.
- Quality tags: Minimal. Strings like
masterpiece, best quality add token noise.
- Negative prompts: Keep short. Guidance-distilled FLUX barely responds to a negative branch.
- Length: Medium (50–100 words).
- Structure:
{subject description}, {setting}, {lighting}, {camera/style}.
SDXL (RealVisXL, Juggernaut XL, DreamShaper XL, etc.)
- Style: Quality tags at the front help significantly. SDXL has two CLIP text encoders trained on quality-rated captions.
- CFG: 7–9. Needs real guidance pressure; below ~6 it looks washed out, above ~10 it over-saturates.
- 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, e.g.
(important:1.3). Keep weights modest (1.1–1.4).
SD 1.5
- Style: Tag-based. Community SD1.5 checkpoints are heavily booru/tag trained.
- CFG: 7–8.
- Quality tags: Essential.
- Length: Shorter (30–80 words). Hard 77-token window; anything past it is truncated.
- Structure:
{quality}, {trigger}, {subject}, {details}, {style tags}.
Wan 2.1 / 2.2 (Video)
- Style: Concise motion descriptions. The sampler animates whatever action you describe.
- 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}.
AnimateDiff
- Style: Same as the base model (SD1.5 or SDXL) for content. Add explicit camera and subject motion keywords.
- Length: Same as the base model.
4. Apply Identity-Method Modifiers
Do not re-describe in text what the method already injects.
- InstantID: Injects identity through a ControlNet-style face branch. DO NOT describe specific facial features. Describe clothing, pose, setting, lighting. Keep CFG at 4–5.
- PuLID: Lighter identity-injection. Can include minimal facial descriptions. CFG 5–7. Use "fidelity" mode to lock closer to reference.
- IP-Adapter FaceID: Encodes reference face into an image-prompt embedding. Describe the style you want, not the face. Use base model's standard CFG.
- FLUX Kontext (Editing): Instruction-conditioned editing model. Describe the EDIT, not the full image. Be specific about what to change and what to preserve.
- Character LoRA: Fine-tuned weights bound to a trigger word. ALWAYS include the trigger word first. Don't describe features the LoRA has learned—focus on what varies: pose, clothing, setting, lighting.
5. Assemble Negative Prompt
- 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:
blurry, low quality, distorted, deformed, ugly, watermark, text
6. Reference Implementation
The rules above are encoded in a strict TypeScript reference implementation. Use this as the single source of truth for prompt assembly.
// Domain types
export type ModelFamily = "flux" | "sdxl" | "sd15" | "wan" | "animatediff";
export type IdentityMethod = "none" | "instantid" | "pulid" | "ipadapter-faceid" | "flux-kontext" | "character-lora";
export interface CfgRange { readonly min: number; readonly max: number; }
export interface CharacterProfile {
readonly name: string;
readonly triggerWord: string | null;
readonly facialFeatures: readonly string[];
readonly hair: string;
}
export interface PromptRequest {
readonly model: ModelFamily;
readonly identity: IdentityMethod;
readonly profile: CharacterProfile;
readonly scene: string;
readonly extraNegatives?: readonly string[];
}
export interface PromptResult {
readonly positive: string;
readonly negative: string;
readonly cfg: CfgRange;
readonly notes: readonly string[];
}
// Rule tables
const MODEL_RULES: Readonly<Record<ModelFamily, ModelRules>> = {
flux: { cfg: { min: 3.5, max: 4 }, qualityTags: [], baseNegative: FLUX_NEGATIVE, maxWords: 100 },
sdxl: { cfg: { min: 7, max: 9 }, qualityTags: ["masterpiece", "best quality", "photorealistic"], baseNegative: UNIVERSAL_NEGATIVE, maxWords: 150 },
sd15: { cfg: { min: 7, max: 8 }, qualityTags: ["masterpiece", "best quality"], baseNegative: UNIVERSAL_NEGATIVE, maxWords: 80 },
wan: { cfg: { min: 5, max: 7 }, qualityTags: ["high quality"], baseNegative: VIDEO_NEGATIVE, maxWords: 50 },
animatediff: { cfg: { min: 7, max: 8 }, qualityTags: ["masterpiece", "best quality"], baseNegative: VIDEO_NEGATIVE, maxWords: 80 },
};
const IDENTITY_RULES: Readonly<Record<IdentityMethod, IdentityRules>> = {
none: { describeFace: true, requireTrigger: false, cfgOverride: null, note: "No identity conditioning — describe the face fully." },
instantid: { describeFace: false, requireTrigger: false, cfgOverride: { min: 4, max: 5 }, note: "Facial features omitted — InstantID provides them." },
pulid: { describeFace: true, requireTrigger: false, cfgOverride: { min: 5, max: 7 }, note: "Minimal facial description allowed — PuLID is robust." },
"ipadapter-faceid": { describeFace: false, requireTrigger: false, cfgOverride: null, note: "Describe style, not face — FaceID holds identity." },
"flux-kontext": { describeFace: false, requireTrigger: false, cfgOverride: { min: 3.5, max: 4 }, note: "Describe the edit instruction, not the full image." },
"character-lora": { describeFace: false, requireTrigger: true, cfgOverride: null, note: "Trigger word first — LoRA conditions later tokens." },
};
// Assembly logic
export function buildPrompt(request: PromptRequest): PromptResult {
// 1. Validate inputs
// 2. Apply quality tags (if model requires)
// 3. Add trigger word (if identity requires)
// 4. Add facial features (if identity allows)
// 5. Add scene description
// 6. Assemble negative prompt
// 7. Return result with CFG and notes
}
Pitfalls
- Hardcoding secrets: Never hardcode secrets in prompt scripts or workflow JSON. Workflow files are routinely shared and embedded in PNG metadata. Read credentials from environment variables.
- Untrusted custom nodes: ComfyUI custom nodes are arbitrary Python that runs in your process on load. An untrusted node pack is a remote-code-execution risk. Prefer well-maintained packs.
- Version drift: ComfyUI's node API and conditioning format change between releases. Pin ComfyUI and custom-node versions. Treat a version bump as something to re-validate.
- Over-describing identity: InstantID, PuLID, and IP-Adapter FaceID approximate a face from a reference. Re-describing facial features in text fights the injected conditioning. Expect to iterate on weight and CFG; do not promise pixel-perfect likeness in a single pass.
- Ignoring token limits: SD1.5 has a hard 77-token window. Exceeding it truncates the prompt mechanically.
- Missing motion keywords: For Wan and AnimateDiff, if you only describe appearance, you get a near-static clip. Motion description is the highest-leverage element.
Verification
Confirm the prompt pair before delivering it to the user:
# Quick verification pass
1. Confirm the target checkpoint and identity method
2. Generate one test image with the drafted positive + negative pair
3. Check quality-tag handling matches the model (no "masterpiece" on FLUX)
4. Verify CFG sits in the recommended range for the method stack
5. Confirm identity holds (features not double-specified for InstantID/FaceID)
6. Adjust weights/CFG and re-run only after the single test image passes
Examples
Example 1: FLUX portrait with InstantID
Positive:
photorealistic portrait, wearing a charcoal wool coat, standing on a rainy city
street at night, neon reflections, cinematic side lighting, shallow depth of field,
shot on Sony A7IV, 85mm lens
Negative:
blurry, low quality, distorted, deformed, ugly, watermark, text
CFG: 4–5 (InstantID range)
Example 2: SDXL with a character LoRA
Positive:
masterpiece, best quality, sage_character, photorealistic portrait of a woman,
detailed skin texture with freckles, emerald green eyes, auburn copper hair,
standing on a rooftop at sunset, wind blowing hair, golden hour lighting,
cinematic composition, RAW photo quality, 8k uhd, film grain
Negative:
(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
CFG: 7–9
Example 3: Wan video (talking head)
Positive:
young woman with auburn hair, talking naturally with gentle hand gestures,
seated at a modern desk, soft studio lighting, high quality
Negative:
static, frozen, jerky motion, low quality, blurry, distorted face,
bad anatomy, glitch, artifacts, flickering, jittery, unnatural movement
CFG: 5–7
Related skills
comfyui-character-gen — builds the identity-preserving workflow / node graph that these prompts feed into (InfiniteYou, FLUX Kontext, PuLID, InstantID, IP-Adapter).
1---
2name: comfyui-prompt-engineer
3version: 1.1.1
4description: "Crafts checkpoint-specific ComfyUI prompt pairs (FLUX, SDXL, SD1.5, Wan, AnimateDiff) with matching negatives, CFG ranges, and identity-method rules (InstantID, PuLID, FaceID, Kontext, character LoRA). Use when generating or adapting prompt text for a known model stack. Not for interviewing a vague idea (comfyui-prompt-interview), wiring node-graph JSON (comfyui-workflow-builder), or installing/launching ComfyUI (comfyui)."
5risk: safe
6source: openrouter-deepsearch
7date_added: 2026-06-16
8metadata:
9 source: https://skillmd.com/skills/kayforkind/comfyui-prompt-engineer
10---
11
12## Overview
13Generates optimized prompts tailored to a specific model and identity method. Diffusion checkpoints do not share a prompt language. Each family was trained with a different text encoder on a different caption distribution, so the same sentence lands differently depending on the target. This skill picks the rules that match the checkpoint and the identity-preservation method actually in the stack.
14
15## When to Use
16Activate this skill when the job-to-be-done is **crafting or refining the prompt text** for a ComfyUI generation. Triggers on requests to:
17- Generate a new positive/negative prompt pair for a specific checkpoint (FLUX, SDXL, SD1.5, Wan, AnimateDiff)
18- Refine or debug an existing prompt that is underperforming on a given model
19- Adapt a prompt when switching checkpoints (e.g., an SDXL prompt that needs to run on FLUX)
20- Tailor a prompt to an identity method (InstantID, PuLID, IP-Adapter FaceID, FLUX Kontext, character LoRA)
21- Build a model-appropriate negative prompt
22- Recommend a CFG range that matches the model + identity-method stack
23- Include the correct trigger word and feature handling when a character LoRA is in play
24
25## Prerequisites
26- ComfyUI installed and target checkpoint identified.
27- Character profile available at `projects/{project}/characters/{name}/profile.yaml` (if using a character LoRA or specific identity).
28- Optional: the Examples section below for fuller prompt pairs; the model-rule tables in Procedure for CFG ranges. This skill crafts prompt text for a known checkpoint — it does not interview a vague idea (use `comfyui-prompt-interview` for that).
29
30## Procedure
31
32### 1. Identify Target Model and Identity Method
33Determine the checkpoint family (FLUX, SDXL, SD1.5, Wan, AnimateDiff) and the identity method (InstantID, PuLID, IP-Adapter FaceID, FLUX Kontext, Character LoRA) in use.
34
35### 2. Load Character Profile
36If using a character LoRA or specific identity, read the character profile from `projects/{project}/characters/{name}/profile.yaml` to get the trigger word and canonical feature list. Do not invent details.
37
38### 3. Apply Model-Specific Rules
39
40#### FLUX.1 (dev / schnell / Kontext)
41- **Style**: Natural-language descriptions. FLUX uses a T5-XXL text encoder alongside CLIP and parses sentences rather than comma-separated booru tags.
42- **CFG**: 3.5–4 (very low) for `dev`; `schnell` runs near CFG 1. High CFG over-saturates color and "fries" the image.
43- **Quality tags**: Minimal. Strings like `masterpiece, best quality` add token noise.
44- **Negative prompts**: Keep short. Guidance-distilled FLUX barely responds to a negative branch.
45- **Length**: Medium (50–100 words).
46- **Structure**: `{subject description}, {setting}, {lighting}, {camera/style}`.
47
48#### SDXL (RealVisXL, Juggernaut XL, DreamShaper XL, etc.)
49- **Style**: Quality tags at the front help significantly. SDXL has two CLIP text encoders trained on quality-rated captions.
50- **CFG**: 7–9. Needs real guidance pressure; below ~6 it looks washed out, above ~10 it over-saturates.
51- **Quality tags**: Include `masterpiece, best quality, photorealistic`.
52- **Length**: Medium-long (50–150 words).
53- **Structure**: `{quality tags}, {trigger word}, {subject}, {details}, {setting}, {style}`.
54- **Weighted syntax**: Supported, e.g. `(important:1.3)`. Keep weights modest (1.1–1.4).
55
56#### SD 1.5
57- **Style**: Tag-based. Community SD1.5 checkpoints are heavily booru/tag trained.
58- **CFG**: 7–8.
59- **Quality tags**: Essential.
60- **Length**: Shorter (30–80 words). Hard 77-token window; anything past it is truncated.
61- **Structure**: `{quality}, {trigger}, {subject}, {details}, {style tags}`.
62
63#### Wan 2.1 / 2.2 (Video)
64- **Style**: Concise motion descriptions. The sampler animates whatever action you describe.
65- **CFG**: 5–7.
66- **Quality tags**: Minimal.
67- **Length**: Short (20–50 words).
68- **Focus**: Describe the motion, not just the appearance.
69- **Structure**: `{subject}, {action/motion}, {setting}, {quality}`.
70
71#### AnimateDiff
72- **Style**: Same as the base model (SD1.5 or SDXL) for content. Add explicit camera and subject motion keywords.
73- **Length**: Same as the base model.
74
75### 4. Apply Identity-Method Modifiers
76Do not re-describe in text what the method already injects.
77
78- **InstantID**: Injects identity through a ControlNet-style face branch. DO NOT describe specific facial features. Describe clothing, pose, setting, lighting. Keep CFG at 4–5.
79- **PuLID**: Lighter identity-injection. Can include minimal facial descriptions. CFG 5–7. Use "fidelity" mode to lock closer to reference.
80- **IP-Adapter FaceID**: Encodes reference face into an image-prompt embedding. Describe the style you want, not the face. Use base model's standard CFG.
81- **FLUX Kontext (Editing)**: Instruction-conditioned editing model. Describe the EDIT, not the full image. Be specific about what to change and what to preserve.
82- **Character LoRA**: Fine-tuned weights bound to a trigger word. ALWAYS include the trigger word first. Don't describe features the LoRA has learned—focus on what varies: pose, clothing, setting, lighting.
83
84### 5. Assemble Negative Prompt
85- **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`
86- **Photorealism Negative**: `3d render, cartoon, anime, illustration, painting, drawing, cgi, plastic skin, smooth skin, airbrushed, video game, doll, mannequin, oversaturated, artificial lighting`
87- **Video Negative (Wan / AnimateDiff)**: `static, frozen, jerky motion, low quality, blurry, distorted face, bad anatomy, glitch, artifacts, flickering, jittery, unnatural movement`
88- **FLUX Negative**: `blurry, low quality, distorted, deformed, ugly, watermark, text`
89
90### 6. Reference Implementation
91The rules above are encoded in a strict TypeScript reference implementation. Use this as the single source of truth for prompt assembly.
92
93```typescript
94// Domain types
95export type ModelFamily = "flux" | "sdxl" | "sd15" | "wan" | "animatediff";
96export type IdentityMethod = "none" | "instantid" | "pulid" | "ipadapter-faceid" | "flux-kontext" | "character-lora";
97
98export interface CfgRange { readonly min: number; readonly max: number; }
99export interface CharacterProfile {
100 readonly name: string;
101 readonly triggerWord: string | null;
102 readonly facialFeatures: readonly string[];
103 readonly hair: string;
104}
105export interface PromptRequest {
106 readonly model: ModelFamily;
107 readonly identity: IdentityMethod;
108 readonly profile: CharacterProfile;
109 readonly scene: string;
110 readonly extraNegatives?: readonly string[];
111}
112export interface PromptResult {
113 readonly positive: string;
114 readonly negative: string;
115 readonly cfg: CfgRange;
116 readonly notes: readonly string[];
117}
118
119// Rule tables
120const MODEL_RULES: Readonly<Record<ModelFamily, ModelRules>> = {
121 flux: { cfg: { min: 3.5, max: 4 }, qualityTags: [], baseNegative: FLUX_NEGATIVE, maxWords: 100 },
122 sdxl: { cfg: { min: 7, max: 9 }, qualityTags: ["masterpiece", "best quality", "photorealistic"], baseNegative: UNIVERSAL_NEGATIVE, maxWords: 150 },
123 sd15: { cfg: { min: 7, max: 8 }, qualityTags: ["masterpiece", "best quality"], baseNegative: UNIVERSAL_NEGATIVE, maxWords: 80 },
124 wan: { cfg: { min: 5, max: 7 }, qualityTags: ["high quality"], baseNegative: VIDEO_NEGATIVE, maxWords: 50 },
125 animatediff: { cfg: { min: 7, max: 8 }, qualityTags: ["masterpiece", "best quality"], baseNegative: VIDEO_NEGATIVE, maxWords: 80 },
126};
127
128const IDENTITY_RULES: Readonly<Record<IdentityMethod, IdentityRules>> = {
129 none: { describeFace: true, requireTrigger: false, cfgOverride: null, note: "No identity conditioning — describe the face fully." },
130 instantid: { describeFace: false, requireTrigger: false, cfgOverride: { min: 4, max: 5 }, note: "Facial features omitted — InstantID provides them." },
131 pulid: { describeFace: true, requireTrigger: false, cfgOverride: { min: 5, max: 7 }, note: "Minimal facial description allowed — PuLID is robust." },
132 "ipadapter-faceid": { describeFace: false, requireTrigger: false, cfgOverride: null, note: "Describe style, not face — FaceID holds identity." },
133 "flux-kontext": { describeFace: false, requireTrigger: false, cfgOverride: { min: 3.5, max: 4 }, note: "Describe the edit instruction, not the full image." },
134 "character-lora": { describeFace: false, requireTrigger: true, cfgOverride: null, note: "Trigger word first — LoRA conditions later tokens." },
135};
136
137// Assembly logic
138export function buildPrompt(request: PromptRequest): PromptResult {
139 // 1. Validate inputs
140 // 2. Apply quality tags (if model requires)
141 // 3. Add trigger word (if identity requires)
142 // 4. Add facial features (if identity allows)
143 // 5. Add scene description
144 // 6. Assemble negative prompt
145 // 7. Return result with CFG and notes
146}
147```
148
149## Pitfalls
150- **Hardcoding secrets**: Never hardcode secrets in prompt scripts or workflow JSON. Workflow files are routinely shared and embedded in PNG metadata. Read credentials from environment variables.
151- **Untrusted custom nodes**: ComfyUI custom nodes are arbitrary Python that runs in your process on load. An untrusted node pack is a remote-code-execution risk. Prefer well-maintained packs.
152- **Version drift**: ComfyUI's node API and conditioning format change between releases. Pin ComfyUI and custom-node versions. Treat a version bump as something to re-validate.
153- **Over-describing identity**: InstantID, PuLID, and IP-Adapter FaceID approximate a face from a reference. Re-describing facial features in text fights the injected conditioning. Expect to iterate on weight and CFG; do not promise pixel-perfect likeness in a single pass.
154- **Ignoring token limits**: SD1.5 has a hard 77-token window. Exceeding it truncates the prompt mechanically.
155- **Missing motion keywords**: For Wan and AnimateDiff, if you only describe appearance, you get a near-static clip. Motion description is the highest-leverage element.
156
157## Verification
158Confirm the prompt pair before delivering it to the user:
159- [ ] Target model identified and the matching rule set applied (FLUX / SDXL / SD1.5 / Wan / AnimateDiff).
160- [ ] CFG recommendation matches the model + identity-method stack (identity range wins when present).
161- [ ] Quality tags handled per model — minimal for FLUX/Wan, front-loaded for SDXL/SD1.5.
162- [ ] Trigger word included first when a character LoRA is used.
163- [ ] Facial features omitted when InstantID / IP-Adapter FaceID provides the identity.
164- [ ] A positive prompt and a matching negative prompt were both produced.
165- [ ] Prompt length sits within the model's recommended range.
166
167```powershell
168# Quick verification pass
1691. Confirm the target checkpoint and identity method
1702. Generate one test image with the drafted positive + negative pair
1713. Check quality-tag handling matches the model (no "masterpiece" on FLUX)
1724. Verify CFG sits in the recommended range for the method stack
1735. Confirm identity holds (features not double-specified for InstantID/FaceID)
1746. Adjust weights/CFG and re-run only after the single test image passes
175```
176
177## Examples
178
179### Example 1: FLUX portrait with InstantID
180**Positive:**
181```
182photorealistic portrait, wearing a charcoal wool coat, standing on a rainy city
183street at night, neon reflections, cinematic side lighting, shallow depth of field,
184shot on Sony A7IV, 85mm lens
185```
186**Negative:**
187```
188blurry, low quality, distorted, deformed, ugly, watermark, text
189```
190**CFG:** 4–5 (InstantID range)
191
192### Example 2: SDXL with a character LoRA
193**Positive:**
194```
195masterpiece, best quality, sage_character, photorealistic portrait of a woman,
196detailed skin texture with freckles, emerald green eyes, auburn copper hair,
197standing on a rooftop at sunset, wind blowing hair, golden hour lighting,
198cinematic composition, RAW photo quality, 8k uhd, film grain
199```
200**Negative:**
201```
202(worst quality:1.4), (low quality:1.4), blurry, deformed, bad anatomy,
203bad hands, extra fingers, missing fingers, extra limbs, fused fingers,
204text, watermark, signature, jpeg artifacts, username, error
205```
206**CFG:** 7–9
207
208### Example 3: Wan video (talking head)
209**Positive:**
210```
211young woman with auburn hair, talking naturally with gentle hand gestures,
212seated at a modern desk, soft studio lighting, high quality
213```
214**Negative:**
215```
216static, frozen, jerky motion, low quality, blurry, distorted face,
217bad anatomy, glitch, artifacts, flickering, jittery, unnatural movement
218```
219**CFG:** 5–7
220
221## Related skills
222- `comfyui-character-gen` — builds the identity-preserving workflow / node graph that these prompts feed into (InfiniteYou, FLUX Kontext, PuLID, InstantID, IP-Adapter).