Magnific drops a 38-page image and video prompting handbook: old keyword tricks no longer work on today's frontier models
- Most people still pad prompts with "8K," "ultra HD," and "masterpiece." Magnific's free 38-page handbook says that kit is dead on current models—and actively drags images toward slop.
- Slop is the model's default answer. Every blank you leave—where the frame sits, which instant you catch, what light—gets filled with its most average answer.
- Two paste-ready blocks for any prompt: a cinematic-feel line and a natural skin texture block, plus seven banned words. Ultra sharp, hyper detailed, 8K clarity and friends summon the AI look.
- Any on-image text has to go in quotes, exactly as it should appear, and you brief the layout like a designer brief. Models place type; they don't type.
- Counterintuitively, the strongest video edit isn't inside the video model. Pull frame one, edit it as an image, put it back as the first frame with the original as motion reference. Same shot, same camera move—only the part you named changes.
- Eight hard rules for the video engine: leave the frame and you're gone; track at most three characters across cuts; shoot around mirrors, never into them. Sounds like superstition; each one was written after a render blew up.
What the handbook covers: image, video, and film-grade camera moves
Magnific released a free 38-page prompting guide, The Prompting Handbook. It covers more than stills—it fills in AI video generation and film-style director control over camera moves. The thesis: AI is fully in the reasoning era—you brief models like a director. Drop the dead keywords. Steer with clear, professional visual language.
In practice: the old stack—"8K," "ultra HD," "masterpiece," "extreme detail"—not only fails, it actively pulls the image toward slop. What replaces it is plain-language scene description, like a director briefing a crew or a client briefing a designer: subject, action, light, framing, style, one piece at a time. Models read sentences now. The more concrete you are, the less room they have to freestyle.
Magnific started as an AI upscaler built in late 2023 by two Spaniards, Javi López and Emilio Nicolás—no funding round, no ads. In five months it hit 725,000 registered users; Freepik acquired it in May 2024. On April 28, 2026, Freepik renamed the whole company Magnific. They don't train models; they wire other people's models in so users can pick by job, which is why the handbook's roster is just what's available on the site. Magnific Studios, an internal team of 40-plus, wrote it—people who use these models day-to-day for studios and fashion brands.
What's inside the 38 pages
- What seven image models are good at, and how to assign work
- Six things every image prompt should cover
- How to get on-image text right
- One-line edits—no masks, no full regenerations
- Keeping the same character consistent across dozens of frames
- A field-tested visual vocabulary, plus how to kill the AI look
- Omni architecture: one generation, picture and sound together
- How four video models split the work
- Eight-part structure for video prompts
- What first frames, last frames, and references each control
- Edit pipeline: pull frame one, change it, put it back
- Turn one still into full-scene camera coverage
- Joining multi-shot sequences: five kinds of cut transitions
- Eight hard engine rules, each earned by a broken render
- How to write twelve camera moves into a prompt
- Dialogue, lip sync, and timing to the beat
- A character audition reel that locks the performance
- Appendix: a pin-up cheat sheet
What you can walk away with
Two paste-ready prompt blocks—a cinematic-feel line and a natural skin texture block—plus seven banned words and a go-to negative prompt set.
A job-routing model matrix: seven image models and four video models, so you stop asking one model to do everything.
Two fill-in templates: six segments for stills, eight for video. Follow them and you stop leaving holes.
A video edit pipeline: pull frame one, edit as an image, put it back—same shot, only the part you named changes.
Eight engine rules: leave the frame and you're gone; track at most three characters; don't shoot into mirrors. Fewer render-killing mistakes.
The 38 pages pack twenty-plus real original prompts, each paired with the image it produced—and failed attempts shown as failed attempts. What follows tracks the handbook's own order. English prompts stay English throughout; those are for the model. Translate them and they stop working.
Seven image models, seven jobs: the wrong model beats a perfect prompt
These models aren't interchangeable. Each was built for different work. Picking the right model matters as much as picking the right words—so the routing table sits on page one of chapter one, ahead of every writing tip.
| Model | Best at | Max res |
|---|---|---|
| Nano Banana ProGoogle · Gemini 3 Pro Image | Cinematic / aesthetic work, hardest prompts, hero stills | 4K |
| Nano Banana 2Google · Gemini 3.1 Flash Image | Edits, product & brand, high-volume iteration | 4K |
| Seedream 5.0ByteDance | Posters, packaging, e-comm, style-locked batches (up to 10 refs) | 4K |
| GPT Image 2OpenAI | Complex layouts, infographics, accurate on-image text | 4K |
| Recraft V4.1Recraft | Logos, icons, vector assets, full brand systems | SVG · 2K |
| Flux.2Black Forest Labs | Brand colors to hex, storyboard-list prompts (up to 10 refs) | 2K |
| Krea 2Krea | Texture work from style refs and mood boards | 2K |
Which Nano Banana to pick
Nano Banana 2 landed in February 2026 and can own most day-to-day work: quality close to Pro, much faster, solid text, and the best consistency across multiple references. Default pick for edits, product, and brand. Nano Banana Pro is the November 2025 generation, positioned as the premium option—use it when you need precision, brand-level fidelity, or to polish the hardest prompt. Slower, higher ceiling.
Same price per image, so route by temperament, not cost. Equal pricing removes "save money" from the decision. The only question left: do you need this frame fast, or exact?
The others have their own personalities. Flux.2 was built for control: ten references at once, exact hex colors, structured storyboard lists—great for brand and product pipelines locked end to end. Hand it a vague prompt and you get a polished, mediocre frame. Krea 2 lives at the other pole: built for style, working from style refs and mood boards to dodge the default AI retouch look. The tradeoff is it interprets more than it obeys, and on-image text is weak. Recraft earns its row because it handles layout-heavy posters and art-school illustration—and it's the only one that exports real vector files.
Six things every image prompt should cover: subject, action, scene, light, framing, style
Today's image models read sentences, not keyword piles. Describe a coherent scene and they build it. Reasoning models like Nano Banana Pro and Nano Banana 2, GPT Image 2, and Seedream parse a dense prompt in chunks before they render, so busy frames with many subjects and details still hold up.
The structure is six parts. The diagram below breaks a real cologne-ad prompt so you can see which part of the frame each segment owns:
Join the six parts and you have the full prompt. The result is the image below. Bottle on the right, a crown splash beside it, hard side light across wet slate—everything in the frame was named in the prompt. Nothing was left for the model to invent:
Same structure, different job. Left: treat the prompt like a layout brief—what the copy says, where each element sits, where to leave air. Right: chase a cinematic feel with a few "looks like" words.
35mm film still: a woman in a red coat dashes across a rain-wet crossing at dusk, mid-stride with motion blur, a bus streaking past behind, sodium streetlight and shop-window glow, muted Kodak palette, natural grain, subject on the left third, full frame edge to edge, no letterboxing, no black bars.