To keep garments consistent in an AI fashion film, define the details that must survive each shot, approve reference images before animating them, and review the generated motion against those references. Build the edit around what the clothing communicates. This guide offers a practical production exercise, not a promise of perfect model consistency or a festival judging policy.
Give the garment a role in the film
Start with a sentence about the experience you want to create. “A structured coat gradually feels like shelter” gives you a reason to choose framing, movement and sound. “A beautiful futuristic fashion film” leaves most of those decisions unresolved. Fashion storytelling can be abstract, but the sequence still benefits from a relationship between its images.
For an illustrative exercise, imagine a blue coat with an asymmetric lapel, a silver fastening and a broad shoulder line. The opening shows its silhouette against an empty space; a closer shot reveals the fastening; the ending shows the wearer settling into the coat. This is an invented brief, not a real collection or a festival selection. Its small scope makes continuity problems easier to identify.
Make a reference sheet with three kinds of detail
Record identity details: the silhouette, lapel shape, fastening position and any distinctive pattern. Then record material behaviour: whether the cloth should hang heavily, reflect light sharply or fold softly. Finally, record what is allowed to change: pose, background, lighting or an intentional transformation. This separates a creative variation from a mistake.
Use a full-length view and a few close details that agree with one another. If the front reference has one fastening and the three-quarter view has two, resolve that disagreement before making moving shots. For a real garment, compare against approved photographs or design material you are authorised to use. For an imagined garment, choose an approved design rather than allowing every new image to redefine it.
A useful shot note might read: “Shot 03 — lapel opens towards the wearer’s left; one fastening at the waist; hem remains below the knee; lighting may warm.” Keep directions tied to the wearer or the camera explicitly. “Left” alone can become ambiguous when a subject turns.
Approve the still image, then direct its movement
Runway’s Image to Video Prompting Guide explains that the input image establishes the starting composition, subject, lighting and style, while the text describes movement and timing. It also warns that defects in an input image can become more pronounced in video. The guide is written for Gen-4.5; check the documentation for the model you actually use.
Apply that distinction to the coat exercise. First approve the garment in the still. Then try a small action: “The wearer slowly turns one shoulder towards the camera. The camera remains still.” This is an example to test, not a validated prompt or a guarantee. Assess whether the turn communicates the coat’s shape before adding a camera move, wind or a change of setting.
Keep each test interpretable. If a sleeve changes shape, try reducing the action or choosing a clearer starting pose before rewriting every part of the shot. Record the input, model, prompt and result together. A successful frame at the beginning says little about whether the fastening remains stable halfway through the movement.
Choose a reference method for the problem you have
A composition reference and a garment reference solve different planning problems. Adobe’s Firefly composition reference documentation describes using a video’s edges, depth and arrangement to guide a generation. It also states that this composition control is unavailable when an image keyframe has been added in that workflow.
That makes the production choice worth testing early. If the priority is the exact starting appearance of the coat, begin by evaluating an image-led workflow. If the priority is the staging of a walk through a space, evaluate composition guidance using footage you can use. Neither choice removes the need to check garment identity in the output.
Do not plan a sequence around combining controls until you have checked that combination in your chosen tool. Make one representative shot before producing the full set. A small test can reveal whether the planned movement, framing and clothing detail are achievable together.
Review continuity within shots and across cuts
Watch each candidate at normal speed first. Does the garment remain readable, or does attention move to an accidental change? Then pause at the beginning, middle and end, and inspect the portions around turns, hand contact and occlusion. Compare the lapel, fastening, hem and shoulder line with the reference sheet. These checkpoints help organise the review; they do not replace watching the entire clip.
Next, put adjacent shots together. A close-up can look convincing alone yet contradict the wide shot’s construction. Watch for changes in colour, pattern scale, sleeve length and the side on which a detail appears. Keep a short log with the shot, timecode, discrepancy and proposed action: keep, trim, repair or regenerate.
If a shot only works before a detail changes, see whether that shorter section still serves the edit. If it does not, replace it. A distracting visual effect is rarely rescued by making the cut faster. Conversely, if transformation is the concept, establish when it begins and what it means so that the viewer can read it as an intentional event.
Build rhythm around what viewers need to notice
Assemble a rough sequence before polishing every clip. Give a silhouette enough time to register, then decide whether a detail shot adds information or simply repeats the mood. Test the coat exercise with three beats: approach, detail, release. The durations should follow the material you have, rather than a rule that every generated clip must be used in full.
Review sound separately. A quiet fabric movement, a footstep or a deliberate pause can draw attention to a physical quality. An overly busy soundtrack can conceal that intention. Watch once without sound to check the visual sequence, then listen with the picture hidden to understand what the soundtrack contributes.
Keep a clean master and review each delivery version. When the film includes speech or meaningful audio cues, our guide to captions, descriptions and accessibility checks can help you plan an accessible version.
Prepare a clear project note for the festival context
Explain whether the clothing is a real collection or an imagined design, what was generated, and what people designed, performed, edited or composited. Credit the actual contributions. A concise AI film project breakdown can show one continuity problem and the decision that resolved it.
Milan AI Festival’s Best AI Fashion Film category covers fashion films, lookbooks in motion, runway concepts and designer storytelling created with AI. Real or imagined brands are welcome, and the maximum runtime is 10 minutes. That ceiling is not a target duration: use the length your idea needs. Entry information is now available on Milan AI Festival’s FilmFreeway listing. Check it for current deadlines, fees and rules.
The festival is a VIA Group — Visual Innovation Awards event, confirmed for 15 April 2027 at Cinema Anteo, Milano. Use the official festival page for participation updates. The workflow above is an editorial way to develop a film; it does not add requirements to the competition.
