AI Portfolio Quality Audit Before the Demo Reel

You can have a very impressive AI portfolio in silence, then lose a creative director in thirty seconds when they play your demo reel with their critical eye. What is missing is not the resolution. What is missing is a consistent audit grid applied shot by shot.

A portfolio quality audit is not a taste contest. It is a reading protocol: style consistency, temporal stability, narrative readability, sound design, and the ability to reproduce the level shown. Without this protocol, you present strokes of luck instead of presenting a signature.

In this guide, I give you the method used in demo reel preparation: scored criteria, acceptance thresholds, a prioritized correction plan, and documentation ready to be shown in a meeting. You come out with a version that convinces on a phone screen as well as in a meeting room.

Key concepts for auditing a reel with no complacency

Hands and teeth are lie detectors. If you do not need the hands, put them off-frame or in distant blur. If you need them, plan a tight reframe on the face and leave the hands out of frame. This is not cowardice, it is craft.

Hard light is not a mistake in itself. The mistake is hard light with no direction. Say where the source comes from, its size, its color. North window, green neon as a backlight, tungsten desk lamp. Even if the model simplifies, your viewer brain looks for a lighting hierarchy. With no hierarchy, you get that flat gray that screams AI.

Film references must be lighting references, not subject references. Saying "like Blade Runner" without specifying interior, rain, indirect neon means nothing to a model. Say instead: rain, reflections on the ground, neon in the background, face lit by a soft close source.

The fear of black pushes beginners to lift the shadows up to gray. Keep real black, especially in cinema. Black gives volume. Gray gives the demo.

A clean project folder is worth all the viral workflow promises. Name your files, keep a screenshot of the settings, copy the prompt into a txt. In two weeks, you will thank yourself when a client says "let us go back to version 2".

Camera moves in AI reward modesty. A 5% push-in over ten seconds sells the emotion better than a full orbit that distorts the architecture. If you want dynamism, cut in the edit, do not force the physics in the generation. The edit lies to the camera, the viewer accepts it.

Shadows under the eyes that are too clean give 3D makeup. Add a micro color variation, a bit of red under the blue, a less sharp transition. Humans have layers, not flat layers.

The AI sequence shot is appealing and rarely clean. If you want one, isolate a simple set, a clear action, a slow movement. Otherwise cut into three shots, the viewer will prefer three truths to one lying sequence.

Contrast is not saturation. Pushing the colors to hide a flat image gives a 90s TV ad. First work the curve: blacks that do not fall into mud, highlights that do not burn the skin. When the curve holds, saturation needs much less.

Copyrights and client ethics are not a paragraph at the end. If you work for a brand, document what is generated, what is retouched, what is stock. The technique here does not replace the legal framework. It lives next to it.

Production notes

The final texture is not a last-minute makeup. It is a consistency choice that must stay consistent with the light and the target distribution. Start light, check on two screens, then adjust only if it helps the reading of the main subject.

Depth of field in the prompt: describe the lens and the distance. Anamorphic gives oval bokeh and a soft falloff. Spherical sharp at 50mm gives a rounder, more neutral bokeh. If you specify nothing, the model gives you a "generic" bokeh, often too sharp and too clean.

The reflections in the eyes tell the room. A rectangular catchlight on a "candle only" scene lies. Harmonize the shape of the source with the set. The small consistency details silence the critical brain.

Image upsampling is not always your friend. More steps can crystallize skin textures into stucco. Look for the level where the pores become suggested again rather than drawn. It is often a little before the maximum the interface proudly offers you.

Eyes that are too bright and too blue are a classic AI signal. Lower the saturation on the whites of the eyes, add a micro shadow under the eyelid, avoid the perfect double-symmetric catchlight. The human eye is slightly imperfect, exploit that.

Intermediate resolution is your lab. Work where you can iterate in ten minutes, not in three hours. When a sequence holds, upscaling or regenerating high makes sense. Otherwise you optimize a perfect pixel in a fake scene.

• Flux vs SDXL
• structuring an AI video like a film
• AI videos that look fake
• optimizing your AI workflow

Field workflow: AI portfolio quality audit before the demo reel

Step 1: brief in five lines

A physical, located subject. The dominant emotion in one word. Duration and format. Three lighting references (films, not adjectives). Explicit prohibitions (no neon, no hands in close-up at the start).

Film references must be lighting references, not subject references. Saying "like Blade Runner" without specifying interior, rain, indirect neon means nothing to a model. Say instead: rain, reflections on the ground, neon in the background, face lit by a soft close source.

Palette consistency across several shots is a LUT or a curve, not a hope. Export a reference, stick it on the edge of your screen, match shot by shot. The eye tires fast, the reference does not.

The storyboard, even rough, saves you hours. Three boxes drawn with a pen are worth ten blind prompts. You know where the horizon line is, where the gaze is, where the cut is. The model does not guess your next shot, you have to give it like a frame.

Shadows under the eyes that are too clean give 3D makeup. Add a micro color variation, a bit of red under the blue, a less sharp transition. Humans have layers, not flat layers.

Step 2: locked image pilot

You only move to video with an image that holds at the skin and fabric zoom. PNG export, archived prompt, noted seed.

The rhythm of an AI clip is built in the edit. If you wait for the generation to give you the rhythm, you will be dependent on chance. Generate shots longer than necessary, then cut hard. The hard cut gives the intention. The dissolve gives the parenthesis. Too many dissolves, and you fall back on the demo clip.

The reflections in the eyes tell the room. A rectangular catchlight on a "candle only" scene lies. Harmonize the shape of the source with the set. The small consistency details silence the critical brain.

A tool's limit is not a personal insult. If a model does not hold the hands, work around it. If another does not hold profile faces, change the angle. The professional studio chooses the tool for the task, not the reverse.

Palette consistency across several shots is a LUT or a curve, not a hope. Export a reference, stick it on the edge of your screen, match shot by shot. The eye tires fast, the reference does not.

Step 3: modest video generation

Duration 3 to 5 s, movement 20 to 45%, one action, an almost static camera or a light push. Batch of four, brutal A/B/C sort.

The working files must survive a computer change. Also export a version that stays readable to you in ten years: mp4 h264 for preview, wav for sound, png for references. Technology changes, archives remain.

Depth of field in the prompt: describe the lens and the distance. Anamorphic gives oval bokeh and a soft falloff. Spherical sharp at 50mm gives a rounder, more neutral bokeh. If you specify nothing, the model gives you a "generic" bokeh, often too sharp and too clean.

Seeds are there to reproduce, not to magically improve. If an image is bad, changing the seed at random is playing roulette. Change the prompt, change the light, then lock a seed when you get close to the goal. Note the seed in your session file, like an operator notes a focal length.

Global sharpening is the enemy. If you want sharpness, mask the face and sharpen very little on the fabrics or the distant details. Never on the foreground skin, unless you are deliberately after a 2000s ad look.

Step 4: sound and editing

Immediate room tone. Hard cut rather than an AI dissolve between different geometries. Fine grain, curve before saturation.

The "teal and orange" grade works when the skin stays human. If everything goes orange, the faces burn. Isolate the skin with a soft mask, bring a real blood tone back into the reds. Even in AI, you will often finish in post. Accept the round trip.

The vertical format imposes a different reading. A horizontal wide shot tells the environment. A vertical demands a clear subject, a strong line, few parasitic elements on the edges. If you reframe a horizontal into a vertical without rethinking the composition, you get cut-off heads and hands that enter by surprise.

The rhythm of an AI clip is built in the edit. If you wait for the generation to give you the rhythm, you will be dependent on chance. Generate shots longer than necessary, then cut hard. The hard cut gives the intention. The dissolve gives the parenthesis. Too many dissolves, and you fall back on the demo clip.

The working files must survive a computer change. Also export a version that stays readable to you in ten years: mp4 h264 for preview, wav for sound, png for references. Technology changes, archives remain.

💡 Frank's Cut: if you hesitate between two versions, keep the one that holds on mobile without you explaining why it works. The explanation in a meeting is already a debt.

Scenario A: intimate interior

North window pilot, wool sweater, a single action (opening a letter with no hand close-up). Video 4 s, push 3%. Light rain sound.

Scenario B: dusk exterior

Wet coat pilot, reflections on the ground, the subject stops. Static camera. Post desaturation 8%, fine 35mm grain.

Scenario C: client deliverable

Eight shots max, same LUT, no tool change mid-dialogue scene. One-page PDF: owned debts, AI chain mentioned if the contract requires it.

The partial face mask, a hat, a strand of hair, can help consistency if your tool struggles on the features. It is not cheating, it is styling. Many real films use off-frame for the same reason.

The weather in an exterior scene changes everything. Same street, same actor, fog or low sun, it is not the same emotion. Set the time of day and the weather in the brief, otherwise the model mixes dramatic clouds with midday light.

"Ultra detailed" prompts often contradict themselves. Adding five different styles in the same paragraph is asking the model to cheat. One dominant style, one concession, one prohibition. Three layers, not fifteen.

Troubleshooting: what beginners break

Face that breathes. Movement too strong or pilot too smooth. Lower the amplitude, get the skin texture back.

Color that jumps between shots. Two contradictory prompts or no common grading session.

Fused hands. Close-up + complex gesture. Wider shot or hands off-frame.

Set that ripples. Tracking shot on vertical lines. Static camera.

2005 TV ad look. Saturation and sharpening before the light. Get the source hierarchy back.

All-nighter at 40 attempts. No pivot rule. Twelve attempts max then change one lever.

Prompts that list twenty aesthetic adjectives with no geometry produce wallpapers. Replace half the adjectives with physical data: distance, focal length, camera height, time of day, dominant material.

Shadows under the eyes that are too clean give 3D makeup. Add a micro color variation, a bit of red under the blue, a less sharp transition. Humans have layers, not flat layers.

The historical Instagram square format is not the same as the TikTok vertical. The visual center of gravity rises in vertical. Place the important information in the upper third, otherwise the phone eats it under the viewer's thumb.

A tool's limit is not a personal insult. If a model does not hold the hands, work around it. If another does not hold profile faces, change the angle. The professional studio chooses the tool for the task, not the reverse.

The reflections in the eyes tell the room. A rectangular catchlight on a "candle only" scene lies. Harmonize the shape of the source with the set. The small consistency details silence the critical brain.

Framings that are too centered give a poster, not a scene. Shift the subject, leave space in the direction of the gaze. The rule of thirds is not a law, it is a tool to avoid the default symmetric postcard.

Computer animation and H.264 compression remind us that compression and temporal consistency matter as much as resolution.

FAQ

Foire aux questions

Réponses rapides aux questions les plus fréquentes sur cet article.

Where do I start with the AI portfolio quality audit before the demo reel without getting lost?

+

Global sharpening is the enemy. If you want sharpness, mask the face and sharpen very little on the fabrics or the distant details. Never on the foreground skin, unless you are deliberately after a 2000s ad look.

The background noise of a night scene is never silent. Even "silence" has a hiss. Add a low room tone, then cut in the edit where you want the real void. The contrast between almost nothing and nothing makes the tension.

Dialogue sequences in AI need reaction shots. Even if you have no real actor, think cut, reverse cut, silence. The edit carries the dialogue, not a single shot that talks for thirty seconds.

How much time should I plan for a first credible A shot?

+

A clean project folder is worth all the viral workflow promises. Name your files, keep a screenshot of the settings, copy the prompt into a txt. In two weeks, you will thank yourself when a client says "let us go back to version 2".

The rhythm of an AI clip is built in the edit. If you wait for the generation to give you the rhythm, you will be dependent on chance. Generate shots longer than necessary, then cut hard. The hard cut gives the intention. The dissolve gives the parenthesis. Too many dissolves, and you fall back on the demo clip.

Contrast is not saturation. Pushing the colors to hide a flat image gives a 90s TV ad. First work the curve: blacks that do not fall into mud, highlights that do not burn the skin. When the curve holds, saturation needs much less.

What is the number one mistake on this subject in generative AI?

+

Social compression noise is a second design layer. If you export too clean, the platform adds its own ugliness. Export with a light grain and control of the highlights, you will gain stability after upload. It is not cheating, it is knowing the medium.

The background blur must follow a distance law. If the nose is sharp and the wall behind is blurred like cream while it is fifty centimeters away, the brain screams fake. Describe the camera-subject distance and the subject-background distance, even approximately.

The viewer looks at the eyes first, then the mouth. If the eyes are sharp but the mouth melts, it is over. Prioritize sharpness on the face triangle, let the rest breathe in optical blur. That is also how many real lenses work.

Should I do everything in a single tool?

+

A clean project folder is worth all the viral workflow promises. Name your files, keep a screenshot of the settings, copy the prompt into a txt. In two weeks, you will thank yourself when a client says "let us go back to version 2".

Kitchen or bar moods with a thousand reflections call for cautious angles. If you simplify a row of bottles into a dark wall, you gain credibility. Reduce the complexity when the model shows limits.

Intermediate resolution is your lab. Work where you can iterate in ten minutes, not in three hours. When a sequence holds, upscaling or regenerating high makes sense. Otherwise you optimize a perfect pixel in a fake scene.

How do I validate on mobile before delivering?

+

Seeds are there to reproduce, not to magically improve. If an image is bad, changing the seed at random is playing roulette. Change the prompt, change the light, then lock a seed when you get close to the goal. Note the seed in your session file, like an operator notes a focal length.

The "teal and orange" grade works when the skin stays human. If everything goes orange, the faces burn. Isolate the skin with a soft mask, bring a real blood tone back into the reds. Even in AI, you will often finish in post. Accept the round trip.

Framings that are too centered give a poster, not a scene. Shift the subject, leave space in the direction of the gaze. The rule of thirds is not a law, it is a tool to avoid the default symmetric postcard.

Can I mix AI and real shooting on the same project?

+

Hard light is not a mistake in itself. The mistake is hard light with no direction. Say where the source comes from, its size, its color. North window, green neon as a backlight, tungsten desk lamp. Even if the model simplifies, your viewer brain looks for a lighting hierarchy. With no hierarchy, you get that flat gray that screams AI.

Subtle camera noise, a micro tremor, can save a shot that is too clean. But a pixel dancing on a cheek is an alert. If the tremor modifies the skin, reduce the amplitude or freeze the face and move only the environment. Separate face and set in your motion strategy.

Eyes that are too bright and too blue are a classic AI signal. Lower the saturation on the whites of the eyes, add a micro shadow under the eyelid, avoid the perfect double-symmetric catchlight. The human eye is slightly imperfect, exploit that.

What should I document to find the same result again in two weeks?

+

Image upsampling is not always your friend. More steps can crystallize skin textures into stucco. Look for the level where the pores become suggested again rather than drawn. It is often a little before the maximum the interface proudly offers you.

Depth of field in the prompt: describe the lens and the distance. Anamorphic gives oval bokeh and a soft falloff. Spherical sharp at 50mm gives a rounder, more neutral bokeh. If you specify nothing, the model gives you a "generic" bokeh, often too sharp and too clean.

When you talk about cinema to a model, think physical camera. A 35mm indoors is not the same thing as an 18mm in the same spot. The 35mm brings the face closer without distorting the shoulders. The 18mm stretches the hands toward the camera and turns a simple gesture into a geometric catastrophe. If your character has hands in the foreground, choose a longer focal length or pull the virtual camera back.