An author’s eye.

An AI’s toolkit.

Scroll Fig. 01 · The machine that looks

Creative AI specialist. Ten years behind the camera for The Guardian, the Royal Photographic Society and the sports industry. Now I build the pipelines, LoRAs and agents behind the image.

PeruTravel film · 2024
PeruTravel film · 2024
More of my work
How I work

The eye

Trained in documentary. I know when an image is right, and why.

See the portfolio →

The engine

ComfyUI, LoRA training, APIs, MCP and Python. Built, tested, documented.

See the builds →

The enterprise

Scoping, rollout and training for global teams.

See the numbers →
Fig. 02 · Behind the image

The image is the last node. Here’s the graph.

Three builds, opened up. An agent that runs a full edit, a character trained on my own face, a product that holds every stitch.

01

The video agent

Claude · Magnific MCP · Python

Three desk takes and a voice-over in. A finished launch video out. Claude plans, calls the tools, edits and verifies against the script. Now a reusable skill.

Raw webcam take of Manuela at her desk
3 desk takes + voice-over
Look image: a bright industrial loft
Look · setting & wardrobe
Claudeplans · calls tools · edits
Magnific MCPSpaces · Seedance 2.5 · upscale
Audiovoice isolation · music · SFX
Python + ffmpegmotion graphics · mix
Self-checkstranscript vs script · −14 LUFS
Skillreusable by the team
Final frame: the same take restyled into a sunlit loft
Output67.7 s launch video · 1080p
Open the Space
The Magnific Space Claude wired through the MCP: raw takes and a look image feeding video generator nodes
Magnific Space wired by Claude through the MCP · raw takes + look → video generator
Download the skill 
magnific-talking-head-video/
├─ SKILL.md        the pipeline + rules learned
├─ scripts/
│  ├─ build_template.py   edit + export
│  ├─ ui_cursor_click.py  UI click shots
│  └─ node_hook_render.py hook animation
├─ fonts/          Geist, OFL
└─ assets/         window mask
  1. Kick-off and script
  2. Restyle the takes in Spaces
  3. Audio rules that held up
  4. Motion graphics in Python
  5. Verify, then deliver

Works in Claude with the Magnific MCP connector, ffmpeg and Python. Brand frames not included: bring your own.

Download .skill
02

I Contain

Character consistency · ComfyUI

One face, held from shot to shot. An identity LoRA trained on my own dataset, plus fixed first and last frames for the video model to land on.

Base portrait that sets the style: a woman in a cream blouse before draped fabric
Base image · style
Training dataset: captioned photos of Manuela
Dataset · captioned
LoRA trainingidentity · ai-toolkit
Qwen Image Editmulti-angle camera
First frame: the character facing camera
First frame
Last frame: extreme close-up of the character
Last frame
Wan image passfixes identity drift
Wan 2.1 · FLF+ personality LoRA
OutputStable face, light, presence
Open the real graph
Screenshot of the ComfyUI graph: diffusion model, LoRA loaders, camera control prompt generator, Qwen text encoders and KSampler
ComfyUI · Qwen Image Edit 2509 with camera control, LoRA loaders, KSampler
03

The jersey

Product consistency · ComfyUI

Models lose logos, fabric and stitching. This pipeline starts from product photos only and holds them, from still to motion.

Product photo of a black football jersey, front
Product photo · front
Product photo of the jersey, back
Product photo · back
Generated composition: a black box under a beam of light
Flux · composition
Qwen Image Editmulti-angle · detail locked
Generated close angle of the crest and fabric
Angle · crest
Generated close angle of the badge and weave
Angle · weave
First and last framecontrolled camera move
Editmusic · sound design
OutputLogos, weave, materials intact

Client work is confidential. The scale isn’t.

Fig. 03 · Enterprise, in numbers
7enterprise accounts led
4regions: Europe, the Middle East, Asia-Pacific and the Americas
37seats in the largest single rollout
~700people at the largest agency I work with
13production flows designed for one engagement
10+years making images before the machine

Consumer tech · Fashion · Retail · Advertising · Energy

Engagements

I work with creatives.

One question first: where AI raises the work, and where it shouldn’t touch it.

  1. 01

    Audit

    How your team produces today, and where AI belongs in it.

    Outcome
    • Production map
    • Prioritised opportunities
    • Tooling and cost model
    1–2 weeks
  2. 02

    Build

    Pipelines for imagery, video and brand consistency. Built on your briefs and your tools.

    Outcome
    • Pipelines tested on live briefs
    • Brand rules built in
    • Full documentation
    2–6 weeks
  3. 03

    Train

    From leadership to the people who run the work every day.

    Outcome
    • Hands-on workshops
    • Leadership briefing
    • In-house power users
    Half or full day
Fig. 04 · ContactMadrid · ES / EN

Engagements, talks and advisory.

manuela.henaor@gmail.comEmail
© HERES · Manuela HenaoAn author’s eye. An AI’s toolkit.