AI creative technologist.
Art direction, film & working tools.
Virtual characters. Product worlds. Social campaigns. Released music and film. I connect image, video and language models through directed, reviewable workflows.
Explore selected work New: ORLO · A campaign & the studio behind it ↘
A song about staying with yourself.
A world that makes that feeling physical.
Lyrics became a treatment. Rooms, cloth and a recurring character became its visual language. Generated takes became a complete music film, then purpose-built portrait edits.
Neurohop / experimental bass
GET IT Studio
A directed generative recording: evolving bass, contrasting voices and a hook that carries the story. The approved song remains the film’s sole soundtrack.
22-second excerpt · Complete track on SpotifyI won't leave myself behind
Just to be somebody in their eyes.
Let my hands shake
when I cross that line—
Every step from here is mine.
I won't leave myself behind
Just to be somebody in their eyes.
Let my hands shake
when I cross that line—
The portrait release uses a selected chorus and a dedicated picture edit. Subject, lyric and musical payoff survive the smaller screen.
See the released short ↗Six moments from the released cut. Select a frame to watch how the image, action and song meet.
01 / 00:00 ↗An imposed fit begins to loosen.
02 / 00:20 ↗A room can be commanded, but still feel empty.
03 / 01:01 ↗Private judgment takes on a monumental scale.
04 / 01:27 ↗An inner verdict becomes a physical place.
05 / 02:06 ↗The repeated promise turns into an action.
06 / 02:56 ↗The film ends with vulnerable self-recognition.
GET IT Studio is my independent music project. Still Mine was released on YouTube on 25 September 2026, with its recording linked on Spotify. Music was generated through a directed Suno V6 workflow; this is not a claim of hand-built synthesis or a manually performed recording. The original film production records contain 38 unique moving sources and a 39-edit timeline. Local LTX-2.5 motion, reference frames, selected takes, an unchanged approved song master and separate portrait edits underpin the released result. The portfolio uses lighter web encodes of the original exports.
Rain, wet cloth and a red signal. A synthetic character holds the frame while material, light and a restrained performance give the world its presence.
Frames extracted from the original sequence. Character features, wet surfaces, rain and the red circular motif are the visual anchors. This is an independent stylised motion study; no client commission or commercial outcome is implied.
A fictional Sicilian accessories label. One virtual cast member and one distinctive bag, directed across daylight, the coast and the end of the day.
Build an independent accessories campaign for a design-aware audience: believable daylight, a recognisable person and a product with a clear physical identity.
Curly hair and a natural face. Ecru poplin and charcoal tailoring. Oxblood leather, a sweeping flap, a silver arc clasp and one strap.
Eight selected stills, a four-view character sheet, two portrait spots, a controlled environment revision and designed social exports.
A step and a shift of weight in the afternoon. A glance back at blue hour. Two short portrait spots bring the same campaign into the feed.
Local LTX-2.5 · Reference-led, two-stage motionTake the studio reference into a blue-hour stairwell. Retain the person, outfit, pose and the Arc bag’s defining construction.
The revision was generated from the working Brand Prompt Studio below. Its reference, brief, seed, workflow and result are saved together for review.
See the tool behind the take ↘
Four views. One casting direction. ↗Reference-led image generation supplies the campaign stills. The local FLUX 3 revision changes the setting and light. LTX-2.5 supplies two reviewed motion takes, followed by a channel edit and a designed end card.
Photographic detail, hands, bag construction, strap attachment, contact shadows and temporal drift remain review criteria. Exact typography is added in a separate design layer.
Download the executed revision graph ↓
ORLO is an original fictional portfolio campaign, not a client commission. The person, product and locations are synthetic. The stills were made through reference-led OpenAI image generation; the tool did not expose an exact model ID. The environment revision ran locally through FLUX 3 Image Klein in the real browser application. Each 8-second portrait spot uses a reviewed selection from a 129-frame source and a designed end card. Fine facial and product details can vary during generated motion; reviewed selections do not imply pixel-identical reproduction. No custom model training, campaign spend or commercial performance is claimed.
Download the campaign handoffA reusable style profile, an editable brief, real generation and a visible review decision. A campaign workspace built on my existing Brand Prompt Studio product.
Reuse a saved profile, choose a reference, edit the art direction and set the composition. The input stays attached to the take.
Inspect the output, compare its reference, keep notes, reject a weak take or use a selected image as the next revision source.
Export selected masters, designed channel layouts and the record behind each take. Branding remains in a controlled finishing layer.
This extension uses the existing Next.js app and its saved style-profile system. A backend submits an exported native ComfyUI workflow, follows real job progress and stores the reference, prompt, seed, graph and output. Review decisions survive reloads.
The blue-hour environment revision above was submitted from this browser interface and completed through local FLUX 3. Imported campaign sources keep their actual provider credit.
Inspect the recorded browser-generated take ↗
Actual native workflow / Open full capture ↗Working local prototype. Reference upload, generation, progress, saved versions, review, revision and export are implemented. The portfolio walkthrough uses actual application captures; access to the generation service is local during this review.
An ordinary apartment. A warning on the phone. A way out that brings a man back to a moment he has already lived.
Build a psychological short around a closed causal loop. The ending changes what the first phone call means.
One original cast, one apartment, practical lighting and an ordinary piece of clothing that becomes a clue. The warning is one exact performance, heard from both sides.
58-second short. Generated dialogue and original score. English captions, ten selected motion takes and an executed 41-node ComfyUI workflow.
Original fiction with an AI-generated cast, location and motion. Photographic source frames establish identity, wardrobe and door geography. Local LTX-2.5 full-model sampling is followed by learned 2× latent refinement and native dialogue decoding. The edit repeats the same warning recording, selects performances before expression drift and uses four static phone inserts to clarify the call and time reversal. The wide knocking take is excluded; a close insert carries one decisive contact. The score is an original Suno V6 instrumental through GET IT’s existing integration.
The source warning is one native LTX recording. Filtering changes its perspective; the words, breaths and pauses stay identical. Shot graphs, source frames and selected intervals remain inspectable.
Inspect the actual ComfyUI workflow ↘A fictional portable speaker becomes a complete visual brief: materials, a recognisable object, a campaign world, controlled variations and compositions made for different formats.
A portable idea.
A bigger world.
The same source image, with a specific edit brief. Drag to compare the orange and green variants. Object shape, grille, framing and lighting remain the review criteria.
Reference-guided image edit · OpenAI image generation
New independent campaign study, made for this portfolio. FIELD 01 is a fictional product, not a client commission or a manufactured device. Reference, campaign, macro, colour and format images used OpenAI image generation and reference-led editing. The character campaign below adds local LTX-2.5 motion and a designed social delivery set. Product geometry, handle attachment, grille, dial, wordmark and lighting are the visual review criteria. Image generation is not a guarantee of exact product reproduction; the source files and review samples remain inspectable in the workflow package.
Meet Mara. A fictional AI persona with a consistent visual identity, brought into the FIELD world. One character brief becomes editorial photography, a moving performance and a coordinated social campaign.
Walk. Pause. Listen.
Two selected performances · LTX-2.5 full / 2× refinementAn unretouched-looking casting direction. A short bob, a worn copper jacket and the same product. The casting portrait and the speaker enter as separate references.
Identity reference ↗
Product in context ↗Mara is a fictional adult AI character in an original portfolio campaign. A new casting portrait establishes a mature face, natural asymmetry and skin variation. The product compositions combine that reference with the original FIELD speaker; the setting uses simple harbour backgrounds and soft overcast daylight. Two local LTX-2.5 full-model runs, followed by learned 2× latent refinement, supply the performances. The final eight-second edit uses selected intervals, separate branding, an end card and the GET IT Studio soundtrack. Product geometry, fingertips, facial movement and the chosen cut points are reviewed directly.
I follow new releases, test them against real briefs, and keep what improves control, quality or turnaround. Explore the working stack and the experiments behind it.
The same photographic source guides face, wardrobe and location. The local model generates both the moving performance and the exact warning. The film reuses this recording from the receiver’s and caller’s perspectives.
20-step full-model base · learned 2× refinement · 8-second native dialogue take · spoken words checked with local ASR
Read the exact executed graph ↗A text-directed workroom study: a hand moves across a table, then the character looks up. An alternative motion model from my GET IT experiments, with its source graph retained.
Inspect the source graph ↗Separate text-to-video experiment · No shared identity referenceThe character identity and product design enter together. A new pose, scale relationship and setting turn separate references into an editorial campaign composition.
Explore the character campaign ↗Reference-guided generation and editing · Human selectionGenre, vocal contrast, lyrics and selection guide the recording. The approved master becomes the organising timeline for the film and its shorter edits.
Hear the track & read the excerpt ↗Establish scale and a silent world.
A tangible beacon gives the idea a form.
A change in gaze carries the human beat.
A small station indicator confirms the connection.
An archived local language-model experiment, made before The Warning, turns an earlier brief into structured JSON. Its original output remains inspectable; it is not the screenplay of the current short.
Inspect the original planning draft ↗My evaluation loop: read the release, test a useful task, inspect continuity and control, then decide where it belongs in the workflow.
Full-model motion, learned 2× refinement, reviewed cuts and a reusable native graph.
Actual JSON planning run, checked and edited before production.
Local campaign revisions through Brand Prompt Studio: saved references, full prompts, seeds and reviewed outputs.
Explore the working studio ↗Working examples and dated experiments · Research entries indicate releases being studied · Model notes reviewed 24 October 2026
The creative work has an inspectable path behind it. Actual graphs, small custom nodes, a local API runner, saved settings and review sheets make that path reusable.
The output, through time. ↗Actual decoded frames / custom review-node output
The RELAY kit includes ten selected motion graphs, input frames, native editor export, actual review sheets, a portable local runner, utility code, captions and the frame-addressed edit recipe.
Download RELAY Production Kit Read the executed graph ↗Local ComfyUI / documented third-party model dependenciesVRCampaignSpec validates model-aligned landscape, cinematic, square and portrait sizes, an 8n+1 frame count and a frame rate. It supplies those values to the actual generation graph and records a delivery specification. VRReviewSheet selects evenly spaced frames, including the beginning and end, and assembles a contact sheet from decoded motion. Both nodes are small production utilities written with AI assistance; they do not train a model or automatically decide artistic quality. VRReviewSheet is used in this film’s executed graph. VRCampaignSpec is a separate format utility used by earlier campaign work. Model weights remain separate dependencies.
Content generation is only one step. I designed BrandToSocial around reusable brand direction, structured briefs, editable drafts and a human approval step before export.
Tone, audience, product facts and content pillars become reusable context.
Draft, edit, approve and export remain separate decisions.
Provider choice, rotation and traceability are part of the documented system concept.
Beauty, coffee and energy. Three fictional packaging directions from BrandToSocial’s source assets.
These are captures of my original BrandToSocial interface, running locally with its own example data. The inspected source includes website-to-brand analysis through GPT-4o mini with structured JSON output. Multi-provider orchestration, rotation and wider content operations are documented product concepts; the full backend for those concepts was not verified here. Packaging visuals are original concept assets, not manufactured products or client commissions. The portfolio makes no claim about users, publishing volume or commercial performance.
My UX/UI background connects the generation pipeline to the person using it: editable direction, visible states and human control over the result.
Available for AI creative production, image and motion direction, custom workflows and product-design collaborations with international teams.
Treatment, story, visual language, shot intent and selection.
Still Mine ↗02Reference-led products, material detail and controlled variations.
FIELD 01 ↗03Fictional personas, product integration, social formats and moving campaigns.
Mara × FIELD ↗04Directed generative music, motion, editorial rhythm and released formats.
GET IT Studio ↗05ComfyUI graphs, Python utilities, API execution and reusable handoff.
RELAY production workflow ↗06Clear inputs, editable direction, review states and human decisions.
Working interfaces ↗I’m Vincenzo Romeo. I combine UX/UI design, creative direction and AI-assisted production to take ideas from a brief to finished media and working tools.
Let’s talk Palermo / Working remotely with international teams
One idea.
Native to the feed.
A 4:5 post, a 9:16 story and a three-slide carousel. Each has its own crop, typography and sequence. The reel gives the same campaign a moving voice.
Feed artwork · 1080 × 1350 · Designed PNG export
Download this artwork ↓Get the complete social set ↓Independent concept · Fictional product and persona