the exact prompts and parameters to use for consistent character generation with ai art tools

I want to share the exact prompts and parameters I use to get consistent character generations across different AI art tools. Over the past year I’ve been iterating on a workflow that balances repeatability and creative variation — the kind that lets you build a character library, maintain brand consistency, or create a cast of characters that look like they belong in the same universe. I’ll give you concrete prompt templates, parameter settings for popular tools, and practical tips to reduce randomness while keeping the results visually interesting.

Why consistency matters and what “consistent” means

When I say consistent, I mean predictable core features: face shape, proportions, key costume elements, color palette, and overall visual style. You still want options — different poses, expressions, or lighting — but the character should be immediately recognizable as the same person. Getting there requires a mix of prompt discipline, reference assets, and model-specific knobs like seeds and fine-tuned weights.

Core strategy — the five pillars I rely on

Clear character sheet language: a short, repeatable description that defines the essentials (age, build, distinct features, wardrobe, color scheme).

Style anchor: an explicit style reference or model token (e.g., “in the style of [artist]” or a model fine-tune/LoRA)

Determinism controls: fixed seed, controlled sampler, and moderate diffusion steps

Reference images: consistent headshot and full-body images used with image-to-image or ControlNet

Negative prompts & constraints: to avoid unwanted artifacts and keep outputs focused

Universal prompt template

Use this as the base and adapt per tool:

Core character sheet: “Character: [Name], [age range], [ethnicity/skin tone], [height/build], distinct features: [scar, freckle, eye color, hairstyle], wardrobe: [garment styles, color palette], personality tags: [stoic, energetic, cunning], lighting: [soft rim light / cinematic], pose: [three-quarter, dynamic action], camera: [35mm portrait / full-body wide], medium/style: [digital painting / cel-shaded / photorealistic], mood: [warm / melancholy].”

Style anchor and variants: “Style: [Studio Ghibli-esque] + [cinematic color grading], reference: [link or uploaded image]. Avoid: [list of unwanted styles or elements].”

Negative prompt: “low quality, deformed hands, extra limbs, text, watermark, blurry, oversaturated skin”.

Exact prompts for specific tools

Below are example prompts and parameter sets I use. Replace bracketed bits with your character details and references.

Stable Diffusion (Automatic1111 / local)

Prompt (positive):

“Character: Amaya, 25–30, East Asian, slim athletic build, short asymmetric bob hair, teal eyes, small silver earring, scar across left eyebrow, outfit: cropped leather jacket (muted teal), high-waist trousers (charcoal), boots; personality: confident yet playful; lighting: soft rim light, cinematic; pose: three-quarter view, slight smile; medium: digital painting, sharp details, painterly brushstrokes; color palette: teal / charcoal / warm skin tones; reference: [image1.jpg]”

Negative prompt:

“bad anatomy, mutated hands, extra limb, lowres, watermark, text, deformed face, over-saturated colors”

Parameter recommendations:

Sampler	Euler a / DPM++ 2M Karras
Steps	20–35 (30 is a sweet spot)
CFG Scale (guidance)	6.5–8.5
Seed	Fixed integer (e.g., 123456789) for reproducibility
Sampler ETA	0.0 or default
Width / Height	512×768 for portrait, 768×1024 for higher detail
Face restoration	GFPGAN or CodeFormer (use with caution; can alter identity)

Notes: Fix a specific seed when you want exact replication. If you want variants, change the seed slightly (e.g., +1, +2) or toggle the “variations” feature and keep the seed locked for consistent base features.

Midjourney

Prompt (single line):

“Amaya, 25–30, East Asian, short asymmetric bob, teal eyes, silver earring, scar on left eyebrow, cropped teal leather jacket, high-waist charcoal trousers, three-quarter pose, soft rim lighting, cinematic digital painting, painterly, warm palette —v 6 --ar 2:3 --stylize 50 --quality 2 --seed 123456”

Parameter breakdown:

--v: model version (use latest stable)

--ar: aspect ratio (2:3 for portrait)

--stylize: 50–100 for controlled stylization (lower = more literal)

--quality: 1–2 (higher = slower, more detail)

--seed: fixed for consistency

Tips: Use the “/blend” or image reference tokens to anchor appearance. Midjourney is more black-boxed, so strong verbal anchors + a reference image give the best consistency.

DALL·E / DALL·E 3

Prompt:

“Photorealistic portrait of Amaya, 25–30, East Asian, short asymmetric bob, teal eyes, small silver earring, scar near left eyebrow, wearing a cropped teal leather jacket and charcoal high-waist trousers, three-quarter pose, soft cinematic rim lighting, shallow depth of field, warm color grading, high detail. Reference image: [upload headshot].”

Parameters: DALL·E has fewer low-level knobs; rely on explicit language and uploaded references. Use the inpainting/edits tool to refine specific facial features across outputs.

Leverage fine-tunes, LoRAs and ControlNet

Consistency jumps dramatically when you use:

LoRA or custom fine-tuned weights trained on your character images (10–50 images often suffice for a recognisable identity).

ControlNet for pose and compositional control — feed a consistent pose skeleton or reference line art to keep silhouette and proportions stable.

Image-to-image with a fixed reference headshot at low denoising strength (0.2–0.45) to preserve identity but allow stylistic reinterpretation.

Reference images and character sheets

Create a small reference pack:

One clean neutral headshot (frontal)

One three-quarter portrait

Full-body turnaround or key costume shots

Color swatches for skin, hair, key clothing

Use these as uploads in your prompt or as anchors for ControlNet / img2img. I keep a “character sheet” file with the prompt template and reference links to paste into any tool.

Negative prompts and artifact control

Always include a targeted negative prompt: hands are common failure points (“deformed hands, extra fingers”), as are teeth and text artifacts. For stylized work, add “photorealistic” to your negative list if you want to avoid realism, or conversely avoid “cartoonish” if you want detail. Tune the negative list as you iterate.

Practical assembly: a copy-paste-ready example

Stable Diffusion example (fill in references):

Positive prompt:

“Character: Amaya, 25–30, East Asian, slim athletic build, short asymmetric bob (deep teal), teal eyes, small silver hoop earring, faint scar on left eyebrow, outfit: cropped teal leather jacket, charcoal high-waist trousers, knee boots; mood: confident playful; lighting: soft rim + warm key light; pose: three-quarter view, slight smile; style: detailed digital painting, painterly brushstrokes, cinematic color grading; color palette: teal / charcoal / warm skin tones; reference: [imgA.jpg], [imgB.jpg]”

Negative prompt:

“lowres, deformed hands, extra limb, watermark, text, badly drawn face, mutated anatomy, oversaturated skin”

Parameters:

Sampler: DPM++ 2M Karras

Steps: 30

CFG: 7.5

Seed: 123456789

Width x Height: 768 x 1024

Img2img denoising strength (if using): 0.32

Iteration workflow I use

1) Start with a fixed seed + core prompt to get a baseline image.

2) Save the result, note any features that drift (nose shape, ear position).

3) Add or tighten prompt details (e.g., “high bridge nose, narrow jawline”) or use a reference crop.

4) If necessary, train a small LoRA on 10–30 curated images, then re-run with the LoRA weight at 0.6–0.85.

5) Use ControlNet for pose consistency when generating full-body shots.

One last practical note: don’t expect perfect identity from a single prompt across wildly different styles. Your job is to lock down the essentials (shape, color, defining marks) and then allow style tokens or model versions to change the rendering. With seeds, references, and a small LoRA, you’ll get consistent characters that still feel alive.