Nobody publishes this honestly, so here it is up front. Two subscriptions, and neither free tier works.
What that means in money. On Higgsfield Plus, 1,200 credits costs $59, so a credit is about 5 cents. A 15 second UGC ad at 1080p is 270 credits, so roughly $13 and 22% of your month. A 30 second cinematic single take is 540 credits, roughly $27 and 45% of your month. Those are Higgsfield's own published figures from their own runs, not estimates.
If it connects but stays silent, start a brand new conversation and name Higgsfield explicitly in your first message. If you get "Authorization with the MCP server failed", go to Settings, Plugins, Higgsfield, and reconnect. Results land in your Assets in about 60 seconds.
These four run entirely inside a ChatGPT conversation. No desktop control, no other software. Start here.
The chain is: your product photo goes in, GPT Image 2 makes the hero plate, then Seedance 2.5 puts it in motion. A product shot at 4K costs 48 credits, then the video on top.
Two instructions do all the heavy lifting here, and both come from Higgsfield's own published prompt. The first stops your packaging text turning to mush, which is the single most common way AI product video gives itself away.
Reproduce the product and ALL its lettering EXACTLY, zero typos, crisp, no warping; the large lettering must stay perfectly readable, the smallest print stays correct or naturally soft-focused, NEVER garbled. The ONLY readable text on screen.The second is the one that stops it looking like a cut-out pasted onto a background. Flat studio lighting on your original product shot will fight whatever room the model generates, and your eye reads that instantly even if you cannot name it.
RELIGHT (CRITICAL): discard the flat studio lighting of the source images entirely. Relight the subject AND the product natively to the generated room, matching its direction, colour temperature and falloff. Both must look physically filmed in this room, not composited.Your presenter can come from three places, and Higgsfield are refreshingly blunt that the third is easiest: generate a new character with Soul 2.0 (0.5 credits), reuse one you saved, or upload a photo of a real person. Their words: "Skip the character sheet if you plan to use a photo of a real person: it works the same way."
A 15 second vertical UGC ad at 1080p runs 270 credits. Speech is generated natively inside Seedance 2.5 through a dialogue block in your prompt, so there is no separate lip sync pass.
Reproduce this exact person, face, skin details, hair, outfit, EXACTLY as shown, nothing invented. Identity only, NOT lighting.Then the part that separates a UGC ad that converts from one that reads as an ad. Perfect is the tell. You have to write the imperfections in deliberately, the same way you would direct a real person to stop performing.
Keep it imperfect on purpose. The propped phone settles a couple of degrees crooked and she never fixes it. Her laugh has a tiny snort. One word gets half-swallowed and she carries on. The framing is slightly off-centre. Natural room tone, no music, no studio polish.This is the skill that matters most if you run paid. It takes one finished ad and makes independent variations: new talent, new wardrobe, new setting, while the product, the motion, the framing, the cuts and the timing stay identical. Source video must be 4 to 30 seconds, and the runtime is about 12 minutes.
Turn this green handbag ad into three campaign-ready variations. Refresh the talent, wardrobe, and upscale home setting while keeping the bag design, product-focused motion, framing, timing, and audio consistent.The variable you swap is up to you. These are the seven canonical versions, straight from Higgsfield's own skill page, and they follow one pattern: name the thing to change, then list everything to preserve.
Use Ad Multiplier to replace the person's outfit in this ad while preserving its motion, framing, cuts, timing, aspect ratio, audio, captions, and all other untargeted text
Use Ad Multiplier to replace the person in this ad while preserving its motion, framing, cuts, timing, aspect ratio, audio, captions, and all other untargeted text. If no person reference is supplied, generate an adult
Use Ad Multiplier to move this ad to a new location while preserving its people, motion, framing, cuts, timing, aspect ratio, audio, captions, and all other untargeted text
Use Ad Multiplier to replace the featured product in this ad while preserving its people, motion, framing, cuts, timing, aspect ratio, audio, captions, and all other untargeted textThis is the most useful thing in the entire guide, so read this bit even if you skip the rest.
There are two completely different ways to make motion graphics here, and almost every complaint you have ever read about AI mangling text applies only to the first one.
The code path gives you things the diffusion path structurally cannot: your exact hex codes, your actual font files with tracking and leading, and real easing curves with adjustable duration and delay. You upload your own SVG logo rather than asking a model to draw one.
One thing to plan for either way: motion graphics come out silent. Higgsfield's own showcase pieces are labelled "Muted" while their UGC examples are labelled "Audio". Music and sound design are always a separate pass.
Four routes, most controlled first: the code engine above, After Effects through the bridge, Marketing Studio's Hyper Motion presets, or a Cinema Studio prompt with your logo attached as the first image. Here is the last one, verbatim, because it is the fastest to try.
Design a logo animation for @Image1. Specific requirements: The logo's appearance from nothing must be creative and well-designed, employing techniques such as particle aggregation, brushstrokes, light sweeping, and geometric decomposition and recombination. The logo's appearance should feature smooth, undulating curves and elastic animation effects, with a strong overall rhythm and professional commercial advertising standards. Light effects sweeping across the logo's surface should create highlights and a metallic texture. The background should be clean and simple, highlighting the logo itself. The color scheme should be completely consistent with the original logo in @Image1. At the end of the animation, the logo should remain stable in the center of the screen with a slight breathing-like looping motion effect.Two craft rules that decide whether it survives. Keep the camera completely still, because 3D rotation invites the model to invent depth and that is when lettering warps. And animate the light, not the mark: a sweep travelling across a static logo is far safer than the logo itself moving through space.
The most finished, least fiddly thing on the platform, and the best starting point if you want a faceless channel. There are 22 built-in styles, so you pick one rather than describing one: Editorial Motion Graphics, Stickman Cartoon, Watercolor Chronicle, Fairy Tale and Myth, Paper Diorama, Pastel Flat 2D, Colorful 3D, Hand Drawn, Poster Vector, Whiteboard Doodle, 3D Papercraft, Mixed Media, Low Poly, Pixel Art, Claymotion, Isometric Flat Vector, Studio 3D, Fluffy Toy, Paper collage and more.
Generate an 8-minute 16:9 editorial motion graphics faceless video about how the Mona Lisa became the world's most famous painting. Show the original commission, Leonardo da Vinci refining the portrait over several years, his translucent glazing technique, and the 1911 theft that drew crowds to the empty museum wall. Use layered collage textures, a restrained burgundy palette, readable typography, and confident documentary pacing.Notice the structure: duration, aspect ratio, style name, then the actual beats in order, then the look. Vague subjects produce vague videos. Listing the beats is what gets you a script rather than a slideshow.
Genjutsu does two distinct things and it is worth knowing which you want. Motion Transfer keeps the movement, the camera and the timing, and rebuilds the scene from your reference images. Object Swap changes one character, outfit, product or location and preserves everything else in the original shot. Input is one video up to 30 seconds plus reference photos, output up to 1080p.
Move the whole fight into this pastel valley by the turquoise stream
Replace the defender with this girl, keeping her watch on her wristWorth being straight about the ceiling here. Shot length is the real constraint. Most models cap at 4 to 15 seconds and only Seedance 2.5 reaches 30, so every story beat longer than that is a cut, not a shot. There is also no rigging or keyframing in the video path, so you cannot direct a performance the way you would direct an actor.
The two most useful lessons published by anyone who has actually finished an AI short, from Higgsfield's own production notes:
And the character sheet technique that makes faces hold up, which is genuinely non-obvious: build three panels, a full body from the front, a full body from the back, and a large close portrait in three-quarter view. Then remove the heads from the full-body panels. On a wide panel the face is small and soft, and that is exactly the face the model will copy into a wide shot, badly. Delete it and there is only one place left to take a face from. Make two close-ups, smiling and not, or the model invents the teeth.
The full skill list on the ChatGPT surface, so you know what you can call by name: Ad Multiplier, AI Host Video, Faceless Video, Higgsfield Presets, Motion Craft, Narrator, Product Photographer, Subtitles, Thumbnail Generation, UGC Product Video, UGC Review Video, UGC Try-On Video, UGC Tutorial Video, UGC Unboxing Video, UGC Website Video, Video Editing, Website Builder, YouTube Script.
This is the part of the video where it edits video and colour grades. It is not the plugin doing that. It is a separate ChatGPT feature called Computer Use, where Astra sees your screen and operates the app. Here is the real setup, because the permission is not where people expect.
The demo everyone shared graded a clip in 4 minutes 19 seconds. Two things about it are worth correcting, because the marketing is looser than the footage. Higgsfield did not do the grading. What is on screen is Resolve's own Color page and its own colour wheels, driven by Astra. And it ran on free Resolve, not Studio.
Set up before you prompt: put both the clip and your reference still into the media pool, then drag both onto the timeline so they sit side by side in the Color page. In the original demo the timeline was empty and Astra had to fix that itself, which cost it a minute.
Go into DaVinci Resolve, open the TEST project, and open Video and Image on the Color page. Color-grade every frame of Video to match Image
Use Image as the primary reference for lighting and color, and grade the footage as a professional colorist would.What you should get back is a two node tree, base correction on node one and the highlight and contrast work isolated on node two. That separation is the point. It means you can adjust the look afterwards without unpicking the whole grade.
For reference, the actual values it landed on: contrast 1.120, gain split red 1.05 and blue 0.92 (that is the warming move), highlights minus 15, saturation 53. Useful as a sanity check that yours is in the right territory.
This one produced 45 separate compositions from a reference video, each element in its own comp, matched to the original timing. Before you start, turn on After Effects, Settings, Scripting and Expressions, Allow Scripts To Write Files And Access Network. It is off by default and the run will stall silently without it.
In the open After Effects project there's a video called Motion_Design and his composition.
Study this video and recreate all the motion design elements individually, each in its own composition. When you're done, create a single composition and place all the elements into it
ignore the person in the videoNotice how short that is, and that it contains no technical instruction about comps, alpha channels or expressions. That is the actual lesson from both of these builds: the specificity came from the model, not from the prompt. The one deliberate exclusion, "ignore the person in the video", is there for a reason. Without it, it will try to rebuild live action too.
Three ways to check you got what was promised, because "editable, transparent background" is easy to claim and easy to fake:
This is the Tesla one, and the anatomy one. Both are open source, and the honest mechanic is better than the myth. Astra did not model 334 car parts from a prompt. It built a polished, interactive Three.js front end around a 3D model that already existed, and the "334 pieces" are the disconnected mesh islands inside that model, split apart by a few lines of Blender Python.
That is genuinely good news, because it means you can do it today, and the source models are free.
Here is a prompt for this exact build shape, published by someone who reproduced it independently. The important thing about it is that it is a shape, not a subject. Get the parts, make each one clickable, explain it, let me take it apart. Swap in an engine, a building, a watch, the solar system.
Build me a fully interactive 3D [SUBJECT] tool. Use real reference geometry where it exists, and use the Higgsfield plugin to generate any visuals you need.
Requirements:
- Every part is a separate clickable object, not one solid mesh.
- Clicking any part isolates it, shows it in 3D, and gives a short plain-english explanation of what it does.
- I can orbit, pan and zoom the whole model freely.
- A sidebar of systems I can toggle on and off independently.
- A slider that pulls the entire model apart into an exploded view and reassembles it.
- A search box so I can type a part name and jump straight to it.
- Dark interface, clean type, no clutter.
Build it, run it, and show me the working result.To split a model into clickable parts, this is the whole trick, run inside Blender:
import bpy
bpy.ops.object.mode_set(mode='EDIT')
bpy.ops.mesh.select_all(action='SELECT')
bpy.ops.mesh.separate(type='LOOSE')To put it online, the easiest route is to say "Use @Sites to deploy this and give me a link", which publishes it on a chatgpt.site URL. Vercel works too, and it is what the original builds used. One warning if you use GitHub Pages: there is a hard 100 MB per file limit and these models get big.
The 45 minute game everyone shared was not one prompt. The person who made it published his whole process as an open source skill, and his follow up is the honest one: "Despite what influencers say, Astra can't one-shot AAA quality Steam-ready games yet. You have to iterate a bit." That build took most of a night and 35 to 50% of a $200 plan's weekly quota.
The reusable version of his method is a loop, and it is the most useful idea in this entire guide. Instead of describing what you want, you make the AI generate a target screenshot first, then build toward it, then have a second AI compare the two and say what is wrong.
Build me a graphics demo: isometric camera, voxel-ish art style with realistic shading and reflective wet floors, a character in an interesting scene. Fantasy setting. Three.js in browser, over 60fps. Don't download assets. Time limit of 1 hour. Controls: click to move the character, camera lazy-follows; drag to rotate camera; scroll to zoom in/out. No gameplay for now. World should feel alive: motion, animations, subtle environmental behaviors. Area around player should look expansive, but only allow movement in a limited space. No need to confirm the art with me or ask questions, just go!The single most important rule in his whole method, and it is counterintuitive: when generating the target image, never use the words "concept art". Ask for a real in-engine screenshot. You are trying to match it pixel for pixel, so if you ask for an artist's impression you get something unmatchable.
And the rubric his critic uses, which you can paste into any visual build to get feedback that is actually actionable:
Score the current screenshot against the target out of 10.
Composition (0-3): camera, framing, layout, position and scale of all major components.
Lighting (0-3): colour palette, exposure, shadows, contrast, atmosphere, reflections, glows. Not too dark or too light versus target.
Materials (0-3): every surface, textures, roughness, translucency, wetness. Not blocky, plasticky, smooth or fake unless the target is.
Details (0-1): not a single pixel should be different.
Avoid non-actionable feedback like "this tree looks fake". Name exactly what is giving that impression and how to fix it. If the score has not improved by a full point in two rounds, stop tweaking and rebuild the scene.The office one. Three prompts, one photo, and it is not what the headline implies. Higgsfield did not build the 3D. Astra built the space out of native Roblox parts from a single 2D photograph, through a connection to Roblox Studio. Higgsfield supplied the textures, five of them: carpet, curtains, desktops, oak trim and acoustic felt.
The three prompts, in order, with real timings: "now, remove everything, and recreate our office, and make it playable" with one photo attached (9 minutes 50), then "does it support multiplayer?" (12 seconds), then "add textures using higgsfield mcp" (13 minutes 31).
That is the pattern worth stealing, and it generalises to anything: Astra drives the domain tool for structure, and calls Higgsfield as the asset layer. Roblox Studio for a walkable space, Blender for models, Three.js for the web. Higgsfield makes what goes on the surfaces.
Here is something no guide tells you. No image model reliably produces seamless tiling textures. Edge matching is an unsolved problem for image generation, so if you generate a texture and tile it, you will see the seams.
The way it actually gets solved is not a better prompt. You generate a nearly-seamless image, then run a deterministic script over it that cuts the seam properly and derives the other maps. Higgsfield's own texture pipeline does exactly this, and produces base colour, normal, roughness and height maps. It does not produce metallic or ambient occlusion, so plan for that.
Wiring the result into a Three.js build:
new THREE.MeshStandardMaterial({ map, normalMap, roughnessMap, aoMap })
// three things that catch people out:
map.colorSpace = THREE.SRGBColorSpace; // albedo only, others stay linear
map.wrapS = map.wrapT = THREE.RepeatWrapping;
map.repeat.set(4, 4);Keep texture dimensions to powers of two, and use a texture atlas if you have a lot of them, because every separate texture is another draw call and that is what kills your frame rate.
Higgsfield publishes these, they are just scattered. All at roughly 5 cents a credit, and remember to multiply by three to ten attempts for anything you actually need to be good.
Two things that quietly cost you money. Subscription credits do not roll over, they expire at renewal, and top-ups expire after 90 days. And asking for 2, 3 or 4 variations bills each one separately.
🎨 Higgsfield: higgsfield.ai
🔌 The plugin directory: chatgpt.com/plugins
🖥️ Computer Use, the official docs: learn.chatgpt.com/docs/computer-use
🧊 The Tesla build, open source: github.com/ashemag/model-x-studio
🫀 The anatomy build, open source: github.com/ashemag/human-atlas
🎮 The game loop, as an installable skill: github.com/achimala/dream-loop
📦 Free 3D models to explode: blendkit.com
🫁 The free anatomy dataset: BodyParts3D