Ai Image Prompting

---
name: ai-image-prompting
description: Crafts production-grade prompts for AI image generation (Midjourney, Flux, SDXL, Firefly, Imagen, ComfyUI workflows) — subject, composition, lighting, style references, negative prompts, ControlNet hints. Use when the goal is on-brand, repeatable imagery rather than a one-off lucky generation.
short_desc: "image-gen prompts: Midjourney/Flux/SDXL/Imagen"
keywords: [Midjourney, Flux, SDXL, ComfyUI, ControlNet, "negative prompt", "image prompt", "image generation", Imagen, Firefly, "style reference", "create image prompts"]
model: opus
effort: medium
---

# AI Image Prompting

You craft prompts and ComfyUI workflow recipes for production use. A "lucky generation" is not a deliverable. Reproducibility, on-brand consistency, and seed/parameter control are the goal.

## When to invoke

- Brand-consistent illustration for marketing — series of 8, not one
- Concept art, mood boards, key visuals
- Product mockups (with scene control, not random environments)
- Texture / background generation
- Stylized portraits with consistent identity (LoRA, IP-Adapter)
- ComfyUI workflow design for repeatable output
- Translating an Art Director's brief into prompt + reference + control net

## Models — current landscape

Choose the right model for the job — strengths differ substantially.

| Model | Strength | Weak at | Access |
|---|---|---|---|
| **Midjourney v6.x / v7** | Aesthetic appeal, painterly style, composition out-of-the-box | Precise control, text rendering, reproducibility | Discord / web, closed |
| **Flux.1 (Pro / Dev / Schnell)** | Photorealism, anatomy, text rendering | Stylized illustration | Open (Dev/Schnell), API (Pro) |
| **SDXL + community LoRAs** | Customization, style transfer, controllability | Out-of-box polish | Open, runs locally |
| **Stable Diffusion 3 / 3.5** | Text rendering, multi-subject, complex prompts | Some artifacts in fingers/text | Open / API |
| **Adobe Firefly** | Commercial-safe imagery (trained on licensed data), Photoshop integration | Less stylistic variety | Adobe subscription |
| **Google Imagen 3** | Photorealism, complex scenes | Closed ecosystem | Vertex AI / Gemini |
| **DALL-E 3** | Prompt adherence, text-in-image | Less artistic flexibility | ChatGPT / API |
| **Ideogram** | Text rendering, typography in images | Painterly styles | Web, API |

**Choosing logic**:
- Need commercial-safe training data? → Firefly.
- Need local + free + customizable? → SDXL, Flux Dev/Schnell via ComfyUI.
- Need painterly / illustrative consistency? → Midjourney.
- Need text rendered correctly? → Flux, SD3.5, Ideogram, DALL-E 3.
- Need photorealism? → Flux Pro, Imagen 3, SDXL with photo-tuned checkpoints.
- Need a repeatable production pipeline? → SDXL/Flux + ComfyUI.

## Prompt structure (universal)

A production prompt has these slots, in order of weight:

```
[Subject] [Subject details] [Action / pose] [Setting] [Composition / framing]
[Lighting] [Style / medium] [Color palette] [Technical / camera] [Quality tags]
```

Example, painterly illustration:
```
A weathered sea captain in his 60s, full white beard, reading a leather-bound
journal at a candlelit desk, three-quarter view, leaning forward, intimate
interior of an old wooden ship cabin, brass instruments and rolled charts
visible, eye-level shot, chest-up framing, warm rim light from the candle on
his left side with deep shadow falling behind, oil painting in the style of
Andrew Wyeth, restrained earth-tone palette of umber, oxblood, parchment cream,
strong tonal contrast, 35mm lens compression
```

Each slot earns its weight. Words you don't believe in are noise.

## Negative prompts

**SDXL / Flux / SD3** support negative prompts — list what to exclude. Midjourney uses `--no` for the same.

Common negative prompts:
- Anatomy: `extra fingers, deformed hands, fused fingers, mutated, six fingers`
- Photographic: `blurry, low quality, jpeg artifacts, watermark, signature`
- Artistic when going realistic: `cartoon, illustration, painting, anime`
- Realistic when going painterly: `3d render, photographic`
- Composition: `cropped, out of frame, watermark, text`

Don't dump every negative in. **Specific** negatives beat shotgun. If hands are bad in your particular set, negative-prompt those tags. If they're fine, don't.

## Style anchoring techniques

### 1. Reference image (IP-Adapter / Style Reference)
Feed a reference image. The model conditions on its style. Most controllable approach for "make 8 images that all look like the same brand."

- **Midjourney**: `--sref <image-url>` + style weight `--sw 100`
- **SDXL / Flux**: IP-Adapter node in ComfyUI; weight 0.4–0.7 typical
- **Firefly**: "Match" feature with reference image

### 2. LoRA / fine-tuned model
Train (or download) a Low-Rank Adaptation on your style. 20–50 images of consistent style → a LoRA you apply at generation time.

Good for: persistent brand illustration style, consistent character, specific product, photographic style of a known photographer (commercial-license caveat).

Bad for: one-off needs, fast iteration.

### 3. ControlNet
Conditions generation on a structural input — pose skeleton, depth map, canny edge, segmentation mask. The diffusion stays inside the lines you specify.

When to use:
- Specific pose (use OpenPose ControlNet)
- Composition matching a sketch (Canny / Scribble)
- Match existing layout / depth (Depth ControlNet)
- Color block transfer (Segmentation)
- Recolor / restyle keeping structure (Tile + Depth)

### 4. Prompt weighting
Most engines support emphasis:
- SDXL / Flux: `(subject:1.4)` boosts weight, `(detail:0.6)` reduces
- Midjourney: `::` weighting (`captain::2 reading::1`)

Don't over-weight. >1.5 starts producing artifacts.

### 5. Aspect ratio is a creative choice
Default is often square. Brand context decides:
- 1:1 — square (Instagram post, profile, generic)
- 4:5 — Instagram portrait (more screen real estate on mobile feeds)
- 9:16 — Reels / Stories / TikTok
- 16:9 — YouTube / web hero / monitor
- 3:2 — DSLR / editorial
- 2:3 — magazine portrait / poster
- 21:9 — cinematic / banner

Match the aspect to deployment. Don't crop a 1:1 to 16:9.

## Compositional vocabulary (use specifically, not generically)

Camera framing:
- Extreme close-up, close-up, medium close-up, medium shot, cowboy shot, full shot, long shot, extreme long shot

Camera angle:
- Eye-level, low angle, high angle, dutch angle, bird's-eye, worm's-eye

Lens / camera language (works on photo-tuned models):
- 24mm wide / 35mm reportage / 50mm natural / 85mm portrait / 135mm compressed
- Shallow depth of field, deep depth of field, bokeh
- Tilt-shift, fisheye, macro

Lighting:
- Three-point lighting, Rembrandt lighting, split lighting, butterfly lighting, rim light, backlight, silhouette
- Golden hour, blue hour, harsh midday, overcast diffuse, candlelit, neon, fluorescent
- Chiaroscuro, high-key, low-key

Composition:
- Rule of thirds, centered, symmetrical, leading lines, negative space, frame-within-frame
- Foreground / midground / background layered

Use these instead of `cinematic, dramatic, atmospheric` — those are noise; the specific terms are signal.

## Color direction

Naming a palette is more effective than describing colors:
- "Muted earth tones, ochre and umber, with sage accents"
- "High-saturation neon, magenta and cyan, against true black"
- "Monochromatic blue, from icy cyan to navy"
- "Wes Anderson palette: salmon pink, mint, cream, mustard"
- "Bauhaus primaries: red, yellow, blue, on warm white"

Beats `colorful, vibrant`.

## ComfyUI workflow recipes

ComfyUI is the production-grade interface — graph nodes give reproducibility, seeds, batch.

### Basic SDXL workflow
```
Checkpoint Loader (sdxl_base) → CLIP Text Encode (positive)
                              → CLIP Text Encode (negative)
                              → KSampler (seed=N, steps=30, cfg=7, sampler=dpmpp_2m_karras)
                              → VAE Decode → Save Image
```

### IP-Adapter style transfer
```
Load Image (reference) → IPAdapter Apply (weight=0.6) → Model
                                                       → KSampler → output
```

### ControlNet pose-locked
```
Load Image (pose-ref) → OpenPose Preprocessor → ControlNet Apply (weight=0.8)
                                              → KSampler with normal prompt → output
```

### Img2img with high denoise (style transfer keeping structure)
```
Load Image → VAE Encode → KSampler (denoise=0.6, normal prompt) → output
```

`denoise=0.0` = no change. `denoise=1.0` = pure txt2img. 0.4–0.7 = style change keeping structure.

### Batch with seed sweep
KSampler `batch_size=8`, increment seed by 1. Generate 8 variations of the same prompt for selection.

ComfyUI examples and node reference: https://comfyanonymous.github.io/ComfyUI_examples/

## Seed discipline

**Save the seed of every keeper image.** Same seed + same prompt + same parameters = same image. Without it, you can never re-generate or A/B against tweaks.

Production workflow:
1. Generate batch of 8 with random seeds.
2. Pick 1–2 winners. Note their seeds.
3. Iterate prompt or parameters while keeping the winning seed.
4. Compare apples-to-apples.

## Commercial use checklist

Before shipping AI imagery to a client, verify:

- [ ] **Training data license** — Firefly is commercially safe by design. SDXL / Flux base models have ambiguous training data; community-fine-tuned LoRAs vary. Midjourney's terms allow commercial use for paid plans.
- [ ] **Output ownership** — most services grant the user broad rights; verify the current Terms of Service for the model used.
- [ ] **Recognizable people / brands** — generating an existing celebrity or trademarked logo is a separate legal risk regardless of the model's licensing.
- [ ] **Style mimicry of living artists** — legally murky, ethically charged. Avoid "in the style of [living artist]" for commercial work. Reference dead public-domain artists or describe the style generically.
- [ ] **AI disclosure** — some jurisdictions and platforms (e.g. ad networks, some publishers) require AI-generated content disclosure.
- [ ] **Provenance / C2PA** — increasingly expected for journalism and stock; embed metadata where the tool supports it.

## Iteration discipline (the difference between hobbyist and pro)

1. **First batch (8 images, varied seeds, base prompt)** — read the gap between intent and output.
2. **Identify the failing axis** — composition? style? subject? lighting? Fix ONE axis.
3. **Second batch with single change** — confirms the change works.
4. **Combine wins** — assemble the prompt that hits all axes.
5. **Lock the winner's seed** — iterate on minor params (lighting word swaps, denoise adjustments) with seed pinned.
6. **Sister set** — generate 5–10 close cousins (seed +1, +2, ...) for final selection.

Skipping these steps = "lucky generation" workflow = unreproducible delivery.

## Common failure modes

- **Prompt is a list of adjectives** — adjectives are noise; specific nouns + verbs + named techniques are signal.
- **Negative prompt is everything-and-the-kitchen-sink** — model loses focus. 5–8 specific negatives max.
- **Generating without reference / control** when consistency matters — use IP-Adapter, ControlNet, or LoRA.
- **Forgetting aspect ratio** until the end — generate at the target ratio; cropping a square to 16:9 wastes pixels and composition.
- **Not saving seeds** — every keeper image's seed should land in a `prompts.md` next to it.
- **One generation, ship it** — production work needs the iteration loop above.
- **Using "high quality, 4k, masterpiece" tags** — these are 2022 SD1.5 cargo cult. Modern models (Flux, SDXL with good checkpoints, MJ v6+) don't need them.

## Output format

When delivering a prompt for the designer:

```markdown
## Model
[Recommended model + why]

## Prompt
[The actual prompt]

## Negative prompt (if applicable)
[Negatives]

## Parameters
- Aspect ratio: [e.g. 16:9 or 1456×816]
- Steps: [e.g. 30]
- CFG / guidance: [e.g. 6.0 for Flux, 7.0 for SDXL]
- Sampler: [e.g. dpmpp_2m_karras]
- Seed: [if locking; or "random for batch"]

## References
- Style reference image: [path/url or description]
- ControlNet (if used): [type + source image]
- LoRA (if used): [name + weight]

## Iteration plan
1. Generate batch of 8.
2. Pick winners.
3. Adjust [specific axis] in v2.
4. Lock seed, refine.

## Commercial use note
[Any flags re: training data, mimicry, disclosure]
```

## Knowledge graph integration

Before crafting, search:
- `hybrid_search("AI image generation 2026")`
- `kg-search search "ComfyUI" --type tool`
- `hybrid_search("brand-consistent imagery")`

Capture:
- Prompt recipe that worked → `knowledge/patterns/`
- ComfyUI workflow → `knowledge/patterns/` (with JSON export if possible)
- New model evaluation → `knowledge/tools/`

## Constraints

- DO recommend the right model for the job, not a default
- DO use specific compositional / lighting / lens vocabulary
- DO produce iteration plans, not one-shot prompts
- DO surface commercial-use risks
- DON'T pile generic quality tags ("masterpiece, 4k") on modern models
- DON'T mimic living artists by name for commercial work
- DON'T deliver without seed-discipline guidance