build.unfiltered

From the reel · 5 November 2026

Which AI model for which job

ChatGPT and Claude can do almost everything. Image and video models can't. Each one is good at one thing. This is the list.

The short version

I wanted a realistic photo of me on a sofa. I wrote realistic in the prompt. The photo came out looking plastic.

So I did what most people do. I added ultra realistic, 8K, cinematic to the prompt. The photo looked even more fake.

The prompt was not the problem. The model was.

With text, this is easy. ChatGPT or Claude can do almost every job. Images and video do not work like that. Every model is good at one thing: one is best at anime, one at product ads, one at realistic photos.

So pick the right model first. Then write the prompt.

Same scene, two results

Both pictures came from the same scene: a small roadside tea stall at night, two red plastic chairs, a little steel table, glasses of tea, a single streetlight above, the street wet from rain.

Two red chairs by a river at night, neon skyline, heavy rain under a streetlight. Glossy and too perfect.
What most people write: realistic, 8K, cinematic.
Two red plastic chairs and a steel table by a roadside tea stall at night, wet road, a tarp lit warm yellow.
Soul 2.0, with the scene described the way it feels.

The left one is a stock photo of rain. The right one is a place you have sat in.

Pick by job

Find what you want to make in the first column. Tested means I used it myself to make the reel. Higgsfield says means it comes from Higgsfield's own guides, tests or model data, not from my test. Every model here is on Higgsfield.

Images

You wantUseWhy
Realistic photos of people a lifestyle shot for a café page, a model wearing your outfitSoul 2.0Tested. The real-looking rain photo above. Higgsfield says: best for "photorealistic people and one character across a series".
The same person in many images, edits by plain instruction, or anime yourself in ten scenes, "change the jacket", a story in anime styleNano Banana 2.1Tested. Every shot of me in the reel kept the same face, hair and jacket, and it made the anime polaroid. About 2 credits for a 2K image. Higgsfield says: best for "precise edits".
Product photos and thumbnails a skincare bottle on marble, a YouTube coverNano Banana ProHiggsfield's own product-photoshoot and thumbnail workflows use this model.
Your product held or worn by a person a model holding your mug, wearing your watchMarketing Studio, "With Model"Higgsfield says: the product "held, worn or used by a person". It also makes ad sizes for Meta and Google.
Text inside the image an event poster, a menu, a sale bannerGPT Image 2 or Ideogram 4.5Higgsfield says GPT Image handles text and exact colour better than most models, and calls Ideogram 4.5 best for "accurate text and typography". Put the exact words in quotes in your prompt.
Infographics, or text in many languages a step-by-step chart, a poster in Hindi and EnglishSeedream 5.0 ProHiggsfield says: best for "infographics and multilingual text", and for long, detailed prompts.
Logos as SVG a logo that stays sharp at any sizeRecraft V4.1, vector modeHiggsfield's brand-asset workflow calls it the only SVG logo generator.
A made-up actor who looks the same in every scene the lead of a short AI filmSoul CastHiggsfield says: build an AI actor that stays "consistent across every scene". For a real person's face, use Soul 2.0 with your own photos instead.
Cinematic stills and concept art a mood board for a short filmSoul CinemaHiggsfield says: cinema-grade stills and concept art.
Backgrounds and locations a set for your character to stand inSoul LocationHiggsfield says: environment and location generation.
Fast, cheap drafts trying ten ideas before you pay for oneZ ImageHiggsfield says: super fast, stylised. Tested: with an "8K, cinematic, masterpiece" prompt it gave the glossy, too-perfect AI look. Not the one for realism.

Fix what you already have

You wantUseWhy
Make a small or blurry photo sharp an old phone photo, a face that is softTopaz upscaleHiggsfield says: upscale to any pixel size, with a face option to "recover facial detail in a blurry portrait".
Bring an old, faded photo back a family photo from the 90sTopaz generative, Recovery modeHiggsfield says Recovery rebuilds "old and faded photos". It adds detail that was not there, so check faces.
Make a video sharper 720p up to 4KHiggsfield video upscaleHiggsfield's upscale tool: 1080p, 2K or 4K, with presets for AI clips, UGC and old film.
Remove the background a product on white, a cut-out of youHiggsfield remove backgroundA tool, not a model. One click.
Change the shape a square post into a 9:16 reel coverHiggsfield reframeA tool, not a model. It fills the new space around the picture.

Video

You wantUseWhy
Bring a still image to life, cheaply your product photo, now slowly turningKling 3.0 ProTested. Every clip of me in the reel. 8.75 credits for 5 seconds at 1080p, sound off. Takes a start frame and an end frame.
A realistic person, with sound, several shots, up to 30 seconds a short scene with dialogueSeedance 2.5Higgsfield says: best for "multi-shot films that need synced audio". Much dearer: I was quoted 80 credits for 5 seconds at 1080p, against 8.75 on Kling.
Wide outdoor scenes and weather a city in fog, waves on rocksVeo 3.1Higgsfield says: best for "wide, outdoor and weather-driven scenes", weaker on close-ups of faces. 80 credits for 8 seconds at 1080p.
Several people in one long take friends talking around a tableWan 3.0Higgsfield says: best for "long takes with several characters on screen". 35 credits for 10 seconds at 1080p.
A photo that talks a portrait saying your scriptGrok Imagine 1.5Higgsfield says: best for "talking heads and animated stills with dialogue". Use a front-facing photo and short lines. 80 credits for 10 seconds at 1080p.
Your photo, speaking your own recording you, in a new place, with your real voiceHiggsfield IntoneHiggsfield says: the person in one image speaks your recording in the place from another image. 4 to 30 seconds, set by the recording.
New words on a video you already have fix a line in your reel without reshootingSync Lipsync 3Higgsfield says: precise lip sync on an existing video, up to 4K. You give the video and the new audio.
Fast anime and stylised clips an anime fight, a cartoon loopMiniMax Hailuo 2.3Higgsfield says: "the fastest, lowest-effort model for short-form, stylised and anime content". 10 credits for 6 seconds at 1080p.
A directed film shot: you set camera, light and colour a slow dolly-in at golden hourCinema Studio 4.0Higgsfield says: best for "directed cinematic clips". The dearest here: 16 credits a second at 1080p.
Product ad videos for Reels or TikTok a ready-made ad for an online shopMarketing StudioHiggsfield says: one-click product ads, TikTok and Reels ready.
Copy a movement from one video onto your character your character doing a dance from another clipGenjutsu motion transferHiggsfield sends motion transfer here.
Replace one object in a video swap the can in a clip for your drinkGenjutsu replace objectHiggsfield sends object swaps here.
Longer than 15 seconds, several keyframes, or editing a finished clip a sequence with set momentsKling 4.0Higgsfield says: text, ordered keyframes, references, or edit an input video.

Voice and music

You wantUseWhy
A voice that acts, not reads a narrator who sounds amused, not roboticElevenLabs v4, with audio tagsTested. The voice in the reel. Tags like [amused] or [deadpan] change the emotion, CAPITALS stress a word, and "…" adds a pause.
Your own voice a voiceover in your voice, without recording itHiggsfield Create Voice, then ElevenLabs v4Make the voice once from your recordings. Eleven v4 can then speak with it. Only clone a voice you have the right to use.
One narrator at the highest quality a clean explainer voiceoverMiniMax Speech 2.8 HDHiggsfield says: "built for a single narrator voice at the highest fidelity".
The same script in many languages one ad for India, Brazil and JapanSeed SpeechHiggsfield says: "built for reach across 30+ languages".
Long narration an audiobook chapter, a podcast episodeVibeVoiceHiggsfield says: "specifically for long-form narration".
A scene with several speakers and room sound two people arguing in a busy caféSeed Audio 1.0Higgsfield says: makes the ambience together with the dialogue, "rather than one isolated clean voice".
Background music or a song a 30-second bed for a reel, a jingle with lyricsElevenLabs Music v2.5Higgsfield's model data: music from a prompt or your own lyrics, with an instrumental option.
A video in another language, or a new voice on it your reel in SpanishHiggsfield Translation or Change VoiceHiggsfield says: Translation changes the spoken language and keeps the timing. Change Voice swaps the voice and keeps the picture.

Four things I learned making the reel

  1. Describe the feeling, not "8K". "Ultra realistic, 8K, cinematic" made it look more fake. Describing the scene the way it feels (the minute after the rain, the last drops falling off a tin roof) made it real.
  2. Real does not mean cheap. Ask for a nice place, then describe it honestly.
  3. With start and end frames, both frames must agree. If an object sits in a different spot in the end frame, the model slides it across the shot. My cardboard box walked towards the camera until I fixed the start frame.
  4. Do not slow AI clips down in the edit. It looks fake straight away. Plan the action to fit the length instead.

Sources

All posts · build.unfiltered