File identity
A checksum identifies the exact bytes of a source or output. It helps detect changes and byte-identical duplicates and supports an auditable chain from original to derivative.
IN DEVELOPMENT / AI PORTFOLIO ENGINEERING
A collection.
A likeness.
A living archive.
James is turning his wife’s extensive photo collection into a managed portfolio — exploring how local AI can organise the material and create new work while keeping the person recognisable.
The technical story behind meyasfashion.com ↗: reference selection, traceable assets, controlled generation and human review.
99files inventoried & checksummed
89image files visually reviewed
10videos sampled for review
One supplied archive reviewed on 10 September 2026. This is a bounded reference batch within the larger collection.
THE PROCESS / FROM SOURCE TO PORTFOLIO
Make each decision traceable.
- 01 / COLLECT
Keep the originals
Inventory source files, record checksums and metadata, check decoding and identify exact duplicates. Keep derivatives connected to their source.
- 02 / CURATE
Build a reference set
Review contact sheets. Select facial detail, expressions, full-figure views and movement references for specific uses. Separate edited mood references from identity anchors.
- 03 / MODEL
Control the experiment
Record the reference assets and generation recipe. Compare short takes with matching words, framing and action before increasing the duration.
- 04 / COMPARE
Measure, then look
Inspect sampled detail, motion and audio measurements alongside the actual footage. Human review checks likeness, natural skin tone, hands, speech and continuity.
- 05 / PROMOTE
Keep an approved set
Save ratings and feedback. The comparison workflow admits an example to its approved set only after explicit approval and ratings of at least 4/5 in each review category.
TWO DIFFERENT CHECKS
What does
“fingerprinting” mean?
Likeness consistency
A curated reference set anchors the intended appearance. Review compares facial shape, expressions, skin texture, hair, visible proportions and movement across takes. James is refining this process as part of his continuing AI work.
A file hash cannot establish a person’s identity. The reviewed project evidence does not establish a validated biometric matching system or a measured recognition accuracy.
WHAT IS IMPROVING
A more disciplined feedback loop.
From folders to a usable catalogue
The reviewed batch has an inventory, checksums, contact sheets and reference selections by purpose. Original files remain intact.
From an attractive frame to a reviewed sequence
The process checks corresponding moments and spoken lines, as well as appearance over time. A sharp still alone cannot establish smooth motion or stable likeness.
From a retry to an informed next take
Saved feedback identifies recurring weak categories and informs a bounded prompt change. This review loop does not automatically retrain model weights.
Paris, from picture to movement.
MF-011 is a 15-second AI-generated fashion sequence built from a still reference. Put the picture beside the result, then step through the movement to see what carries across and what the generation adds.
The scene reference
The selected sequence
Follow the turn.
These are frames from the same film, sampled four seconds apart. Select a frame to inspect that moment in the player; it stays paused until you press Play.
The film loads when you play it or select a frame.
How the picture becomes a sequence.
Choose the reference
The Paris street picture becomes the first-frame reference for image-to-video generation. It sets the navy-and-ivory tweed look, the street setting and the full-body composition.
Map the movement
Timed direction asks for front, side and back views while retaining the outfit and framing. This maps the intended action over time; it does not reconstruct a 3D body or environment.
Generate and review
MiniMax H3 turns the still reference and movement direction into a sequence. Review follows likeness, clothing and continuity across the turn. The four frames above are moments from the finished film, not intermediate generation steps.
Upscale, then prepare for the web
Real-ESRGAN enlarges the generated 1152 × 1536 sequence by 4× to a 4608 × 6144 master. This page plays a smaller 720 × 960 copy of the complete selected take, with still posters and no automatic playback.
From a Space to a short film.
I use my MiniMax H3 workflow to make video summaries of Sunday Builders Club Spaces: taking ideas from the conversation into scene direction, references and generated footage.
The same local production setup also supports the Paris picture-to-film study. The subject changes; the connection between source, prompt and selected take stays visible.
How the work comes together.
Shape the summary
The conversation supplies the story. A written brief identifies the ideas to carry into the film and separates what was said from the visual interpretation.
Direct each scene
Reference images and structured prompts set the action, framing, soundscape and music direction for short H3 clips.
Queue the footage
Pinokio and ComfyUI run on the M3 as the control point. The distributed queue delegates generation to two GX10 render workers.
Review the result
Source references, prompts, job IDs and output records support comparison and selection. Generated scenes illustrate the summary; they are not recordings of the people speaking.
The machines have different jobs.
- M3
- Interface, workflows and queue control.
- GX10 workers
- MiniMax H3 video generation.
- Human review
- Meaning, continuity and the selected take.