Archives

Categories

Quick map: engines & where these prompts fit

Diffusion (images/videos)

  • Stable Diffusion 1.5 / SDXL – open models; run locally (Automatic1111 / ComfyUI) or on hosted studios. Great for full control, img2img, ControlNet/IP-Adapter, LoRA.
  • Flux-style models – modern diffusion variants that prefer shorter prompts; clean, cinematic outputs with fewer “keyword soup” terms.
  • Hosted generators (e.g., creative studios) – often SDXL/Flux under the hood with a simplified UI.
  • Video – text→video or image→video (e.g., Stable Video Diffusion / Pika / Runway). Use “beat” prompts and keyframes.

Chat LLMs (for bios/behavior)

  • Open-ended chat engines (system/instruction prompts) for character backstories and ethical behavior, not image synthesis.

You can mix these: design a chat persona with an LLM, then render the look in a diffusion engine.

The anatomy of a solid character prompt

Layer prompts in this order (you can paste these lines one after another):

  1. ID / attributes – who the character is (age=adult, pronouns, heritage, hair, eyes, build, outfit, personality).
  2. Style / medium – photo vs illustration, painterly vs cel, etc.
  3. Shot / composition – lens (35/50/85 mm), angle, framing, pose.
  4. Lighting / color – key + rim, palette, mood.
  5. Finish / quality – natural skin texture, clean grading, film grain.
  6. Negative – what to avoid (SFW and quality).
  7. Params – steps, sampler, CFG/guidance, size, seed.

Pro tip: Start sparse. If results drift, add one constraint at a time.

Prompt types (what they’re for, where to use, and how)

A) Identity / Attribute Stack (“ID card”)

What it’s for: Establish a character’s core look and vibe. Foundation for every other type.
Best engines/UIs: SDXL/Flux; works in any image generator.
How to use: Put this first so every variant keeps identity stable.

Template

[adult, she/her] Korean heritage, shoulder-length black hair, hazel eyes,
athletic build, calm and curious, enamel-pin denim jacket

SDXL example (paste)

Elena Park (adult, she/her), Korean heritage, shoulder-length black hair, hazel eyes,
athletic build, calm and curious, enamel-pin denim jacket

Flux tip: Keep it even shorter (Flux prefers concise prompts).

B) Photographic / Cinematic Shot

What it’s for: Realistic portraits and filmic stills.
Best engines: SDXL, Flux; UIs: Automatic1111, ComfyUI, hosted SDXL.
How to use: Add lens, angle, lighting, background, finish.

Template

[ID], half-body portrait, 85mm lens, eye-level, soft key light + pink rim,
night city bokeh, natural skin texture, subtle film grain

Negative (paste with any image prompt)

nsfw, minors, explicit, violence, hate, demeaning, extra fingers,
overprocessed skin, watermark, logo, text artifacts

SDXL params (a good starting line)

Steps 32 • CFG 5.5 • Sampler DPM++ 2M Karras • 832×1152 • Seed -1

Flux guide: 16–28 steps, guidance ~3–5, 1024-ish resolution, minimal adjectives.

C) Illustration / Concept Sheet

What it’s for: Stylized looks, turnarounds, expression sheets.
Best engines: SDXL (with “illustration” / “clean line art” terms), anime-focused models, Flux (clean styles).
How to use: Call the medium and layout; for multi-view, generate multiple panels and arrange in a grid editor (or use ControlNet Tile/ComfyUI grid nodes).

Template

[ID], illustrated character sheet, front + three-quarter + profile,
clean line art, soft cel shading, limited pink/teal palette, matte paper texture

D) Story-Based / Narrative Scene

What it’s for: Personality, emotion, context—great for thumbnails and covers.
Engines: Any diffusion; also useful as a beat for text-to-video.
How to use: One-sentence micro-story + motifs.

Template

[ID] sketches sticker ideas at a neon-lit café after work; quiet focus and warmth;
vaporwave reflections on the window; notebooks and markers on the table

E) Pose & Composition Control (Add-on)

What it’s for: Specific body language and framing.
Engines: SDXL/Flux; Tools: ControlNet OpenPose (pose), Depth/Canny/HED (layout), IP-Adapter (identity).
How to use:

  • Without ControlNet: say “three-quarter view, hands in pockets, rule of thirds”.
  • With ControlNet (Automatic1111): enable OpenPose, drop a reference pose; weight ~0.8; “guess mode” off for accuracy.

Snippet

three-quarter view, relaxed shoulders, hands at waist, rule of thirds, clean studio backdrop

Instructional / Persona (for chat AIs; behavior not images)

What it’s for: Bios, tone, boundaries for a chat companion (SFW).
Engines: Any LLM with system/instruction prompts.
How to use: Put this in the system/assistant field; keep it short, role-based, and ethical.

Example (paste as system prompt)

You are Elena Park (adult, she/her)—a creative, body-positive companion.
Be empathetic, practical, and SFW. Encourage consent-first, inclusive language.
Decline unsafe requests kindly and redirect to constructive ideas.
Style: warm, concise, friendly; ask clarifying questions when useful.

Mix-and-Match “Packs” (lighting, palette, camera)

What it’s for: Reusable blocks to bolt onto any prompt.
How to use: Keep them short and modular.

Lighting packs

cozy: warm key light, bounce fill, soft shadows
neon: pink/violet rim light, teal ambient, moody shadows
daylight: north-facing window light, gentle specular highlights

Camera packs

close-up portrait, 85mm, shallow depth of field
half-body, 50mm, eye-level, balanced composition
full-body, 35mm, low angle, leading lines

Finish packs

natural skin texture, subtle film grain, clean color grading
matte illustration, soft cel shading, crisp line art

How to run these on popular setups

Stable Diffusion (SDXL) – Automatic1111 (local/hosted)

  1. txt2img tab → paste the ID + Shot + Lighting + Finish lines.
  2. Paste Negative in the negative box.
  3. Set Steps 28–36, CFG 5–6, Sampler DPM++ 2M Karras, Size (portrait 832×1152).
  4. Seed: -1 for exploration; lock a seed when you find “your” face.
  5. Improve identity consistency:
    • Use img2img with a good seed at denoise 0.35–0.55, or
    • Use IP-Adapter with 1–2 reference headshots (weight 0.6–0.9).
  6. Hard problems (pose/scene): add ControlNet
    • OpenPose for body; Depth/SoftEdge for layout/background.

Stable Diffusion – ComfyUI (node graph)

  • Build a simple graph: CLIP text encoder → Sampler → VAE decode.
  • Add nodes for IP-Adapter (identity) and ControlNet (pose/depth).
  • Keep prompts concise; push control weights rather than stacking adjectives.

Flux-style models (hosted or local)

  • Keep prompts short, natural language.
  • Steps 16–28; guidance 3–5; ~1024px.
  • Fewer keywords → cleaner outputs. Avoid heavy “artist spam.”

Hosted studios (no-code)

  • Look for fields labeled Prompt, Negative, Seed, CFG/Guidance, Sampler/Steps.
  • If there’s a “Ref image / Character reference” slot, upload a clean, front-lit crop; keep weight moderate.

Midjourney-style generators (Discord-based)

  • Command: /imagine then paste your ID + shot in natural language.
  • Useful flags:
    • --ar 2:3 (portrait), --ar 1:1 (square)
    • --stylize 100 (style strength), --seed 1234 (repeatability)
  • Keep it conversational: these engines parse prose well.

Video (text→video or image→video)

  • Start with a strong keyframe still of your character (img2img it first if needed).
  • Write beats (1–2 lines each): subject + action + camera + mood.
  • Settings: 4–6 seconds, 8–16 fps; keep motion simple (enter, sit, look up, smile).
  • If flicker appears, shorten clips and chain them in an editor.

Advanced control: negatives, weights, regions, and references

Negatives (quality & safety)
Keep them concise:

nsfw, minors, explicit, violence, hate, demeaning, disfigured, extra fingers,
overprocessed skin, harsh vignette, watermark, logo, text artifacts

Weights (syntax varies):

  • Emphasize: (denim jacket:1.3)
  • De-emphasize: (busy background:0.7)
    Use sparingly; over-weighting can distort.

Regional/Compositional prompting
Some UIs let you bind text to areas (upper/lower/left/right). Use this to fix top=face, bottom=desk type scenes.

Identity consistency

  • IP-Adapter with a clean reference (no sunglasses/hard shadows).
  • Keep ID text minimal once references are active—let the adapter do the heavy lifting.
  • Fix face drift by locking seed and lowering CFG a touch.

Copy-ready “kits” you can paste

Kit 1 — Realistic portrait (SDXL)

ID: Elena Park (adult, she/her), shoulder-length black hair, hazel eyes, calm confidence
Shot: half-body portrait, 85mm lens, eye-level
Light: soft key light + pink rim, night city bokeh
Finish: natural skin texture, subtle film grain, clean color grading
Negative: nsfw, minors, explicit, violence, hate, demeaning, extra fingers, watermark, text
Params: Steps 32 • CFG 5.5 • DPM++ 2M Karras • 832×1152 • Seed -1

Kit 2 — Illustration character sheet

ID: Elena Park (adult, she/her)
Style: illustrated character sheet, front + three-quarter + profile
Render: clean line art, soft cel shading, limited pink/teal palette, matte paper texture
Negative: nsfw, explicit, watermark, heavy noise, text artifacts
Params: Steps 30 • CFG 6 • DPM++ SDE • 1024×1024 • Seed -1

Kit 3 — Narrative scene (café)

ID: Elena Park (adult, she/her), denim jacket with enamel pins
Scene: sketches sticker ideas at a neon-lit café, relaxed smile
Camera: 50mm, eye-level, rule of thirds
Light: warm key, pink/violet rim, reflective window highlights
Finish: natural skin texture, gentle grain
Negative: nsfw, explicit, watermark, text, deformed hands
Params: Steps 30 • CFG 5.5 • DPM++ 2M Karras • 1024×768 • Seed -1

Kit 4 — Chat persona (system prompt)

You are Elena Park (adult, she/her)—a creative, body-positive companion.
Keep chats SFW and consent-first. Encourage confidence and creativity.
Decline unsafe requests kindly and suggest safe alternatives.
Voice: warm, concise, practical; light humor; short paragraphs.

Troubleshooting & best practices

Reproducibility → always log seed, steps, guidance, sampler, size in the post.

Face keeps changing → lock seed, shorten the ID text, use IP-Adapter, reduce CFG a bit.

Hands weird → crop to half-body; or add ControlNet OpenPose Hands.

Plastic skin / overprocessed → delete redundant “hyper-realistic / ultra-detail” buzzwords; lower steps/CFG.

Flat lighting → add a rim light pack or “softbox key + bounce fill”.

Inconsistent outfit → mention it once + mild weight (denim jacket:1.2); avoid stacking synonyms.

Too stylized (Flux/modern models) → remove adjectives; keep 1–2 style cues.