Structured Prompt Formula
Use this order every time. It forces the model to parse logically instead of guessing.
- [Subject] Main focus, character or object, with key details . Age, clothing, expression, specific pose.
- [Style] Artistic or photographic approach . Photorealistic, cinematic, oil painting, product shot.
- [Composition] Framing, angle, placement . Close-up portrait, wide establishing shot, rule of thirds, Dutch angle.
- [Lighting] Time of day, source, quality . Golden hour side lighting, dramatic chiaroscuro, soft studio key light, neon rim lighting.
- [Technical] Camera, lens, film stock, quality markers . Shot on 35mm, 85mm f/1.8, 8K, sharp focus, complicated details.
Model Comparison (April 2026)
| Model | Resolution | Speed | Best For | Gotchas |
|---|---|---|---|---|
| FLUX.2 dev (32B) | Up to 4MP | 5-20s on API, heavier locally | Photorealism, multi-reference editing, complex scenes | 32B needs serious VRAM. Quantized FP8 helps but still 40-60GB range. |
| FLUX.2 klein (4B) | Up to 2MP typical | Sub-second on consumer GPUs (distilled 4-step) | Fast iteration, local runs | Slightly less coherent on very complex prompts than dev. |
| Midjourney V8 Alpha | Native 2K with --hd | ~5x faster than prior, seconds per image | Artistic coherence, style refs, personalization | --hd costs more GPU time. Still occasional anatomy slips on crowded scenes. |
| Grok image gen | High, flexible | Fast in-app | Creative, less censored outputs, bold aesthetics | Tends toward dramatic or stylized. Prompt adherence varies. |
| Ideogram | High res | Fast | Text in images, posters, clean typography | Excels at readable text. Can feel more graphic-design oriented than photographic. |
| DALL-E / GPT Image | 1024-1792px | Fast via ChatGPT | Natural language prompts, safe outputs | Conservative guardrails. Lower max res than FLUX.2. |
How Each Model Interprets Prompts Differently
FLUX.2 reads like a technical spec. It follows the structured formula strictly and handles long prompts well. Physics and material properties land accurately.[1]
Midjourney V8 Alpha responds to aesthetic language. It thrives on artistic references and mood. The --hd flag gives native 2K without upscaling artifacts. Prompt adherence improved but still benefits from concise style terms.[[2]]. Https://updates.midjourney.com/v8-alpha/
Ideogram treats text as a first-class citizen. Put exact copy in quotes and it renders legible typography almost every time. It leans toward clean, designed compositions.
Grok image gen favors bold, expressive results. It handles creative or edgy directions with fewer refusals but can amplify drama in lighting and color.
DALL-E / current GPT Image follows natural sentences best. It understands conversational prompts without heavy engineering. Output stays safe and coherent but rarely pushes technical limits on resolution or physics.
Aspect Ratio and Resolution Cheat Sheet
Use multiples of 16 or 32 for most backends. FLUX.2 handles any ratio but performs best in multiples of 64.
| Aspect | Ratio | Common Use | Pixel Example (FLUX.2 safe) | Notes |
|---|---|---|---|---|
| Square | 1:1 | Icons, portraits | 1024x1024 or 1536x1536 | Safe default |
| Portrait | 2:3 or 9:16 | Characters, social | 896x1344 or 1080x1920 | Good for people |
| Field | 3:2 or 16:9 | Scenes, products | 1344x896 or 1920x1080 | Cinematic |
| Ultrawide | 21:9 | Film stills | 1792x768 | Use sparingly |
| Tall cinematic | 4:5 | Posters | 1024x1280 | Midjourney friendly |
Consistent Character Workflow
- Generate a clean reference sheet first. Use one strong front, side, and 3/4 view in the same outfit and lighting.
- For Midjourney V8: Upload the reference image, add it at the front of new prompts or use character reference features. Backwards compatible with older personalization.
- For FLUX.2: Feed the reference image(s) as input. Dev variant supports up to 6-10 references reliably. Describe the character precisely then vary only pose, setting, or expression.
- For Ideogram: Use style reference or character tools with up to 3 images. Lock colors if needed.
- Iterate in the same tool first. Cross-tool consistency drops unless you export the reference and re-describe heavily.
- Gotcha: Clothing and exact facial features drift fastest. Fix them in the subject section of every prompt. Avoid changing lighting temperature between shots.
Copy-Paste Prompts
These follow the structured formula. Copy the whole block.
1. Cyberpunk street vendor elderly Asian woman selling glowing tech trinkets from a cluttered stall, cyberpunk street market at night, rain-slicked neon reflections, tight medium shot from slightly low angle, volumetric neon lighting with pink and cyan rim light, cinematic color grade, shot on 35mm anamorphic, 8K sharp focus complicated details
2. Product shot for wireless earbuds matte black wireless earbuds floating above polished walnut desk, minimalist studio product photography, clean white background with subtle shadow, three-quarter view, soft diffused key light from top left with crisp edge highlights, commercial advertising style, 85mm lens f/2.8, ultra sharp 4MP resolution
3. Portrait of female engineer mid-30s South Asian female engineer with safety goggles on forehead, focused expression in modern lab, wearing white coat over plaid shirt, eye-level close-up portrait, soft natural window light from side creating gentle shadows, technical documentary style, shot on ARRI Alexa, high detail skin texture
4. Fantasy warrior in forest tall male elf archer with complicated leather armor and recurve bow, ancient misty forest at dawn, dynamic pose drawing arrow, low angle heroic composition, golden volumetric god rays piercing canopy, epic fantasy illustration style, detailed details on foliage and armor, 2K native resolution
5. Modern kitchen interior bright Scandinavian kitchen with white oak cabinets and marble counters, morning sunlight streaming through large windows, clean architectural photography, wide angle establishing shot, soft warm natural lighting with cool shadows, realistic interior design magazine style, 16:9 aspect, photorealistic
6. Vintage robot mechanic retro 1950s style chrome robot repairing a classic car in garage, detailed mechanical arms and tools, three-quarter view, dramatic workshop lighting with sparks and warm tungsten bulbs, pulp sci-fi illustration meets photorealism, high contrast, sharp mechanical details
7. Abstract data visualization portrait young Black male data scientist surrounded by floating holographic charts and code, dark cyber environment, medium shot with shallow depth of field, cool blue and purple volumetric lighting, futuristic tech aesthetic, cinematic, 4MP resolution with bokeh on background elements
8. Quiet mountain cabin at dusk cozy wooden cabin with warm interior lights glowing through windows, surrounded by snow and pine trees at twilight, wide market composition, soft purple and orange sky gradient with subtle aurora, peaceful realistic market photography, long exposure feel, 3:2 ratio
AI Video Tools (Brief, April 2026)
Image-to-video dominates practical workflows. Runway, Kling AI 3.0, Google Veo 3, and Grok Imagine handle motion best right now. Start with a strong FLUX.2 or Midjourney still, then drive the video generator with a short motion prompt describing camera move and action only. Physics still breaks on complex interactions, and Test short clips first.
Use the reference image as anchor and keep the subject description identical to the original image prompt. That's where consistency lives.