AI Videos Look Bad? Here's Why

2026-03-01 · long tutorial / working session · Adil · recording, 9:14 · watch on YouTube

Topic: Video generation & shot direction (also: Prompting & AI assistants, Image generation & editing, Assets, references & consistency, Post-production & finishing) · mode: Tutorial · tags: prompts shared, giveaway, one creator's experience

In this video tutorial Adil names five mistakes that make AI videos look cheap, starting with text-to-video and no start frame, which he calls a slot machine; he says the start frame decides 80% of the result and makes it in Soul 2.0. He then asks for clear prompts (who, where, what, camera move, mood), a Nano Banana Pro character sheet tagged in Cinema Studio, and one primary action per scene. He closes with tips: let an LLM draft the animation prompt and edit it, use motion control, and generate fast action in slow motion, then speed it up.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — Why most AI videos look cheap

★★ Adil says most AI videos look bad not because of the AI but because of how people use it. — 00:00:00
★★★ Adil calls typing a long prompt and hitting generate, text-to-video, the biggest mistake: basically a slot machine with zero control over composition; the fix is to use a start frame. — 00:00:23 · text-to-video, keyframe

“That's text to video. And it's basically a slot machine.”

00:00:31 — Mistake 1: text-to-video without a start frame

★★ Adil explains that a start frame is where the video begins, the image being animated: without one, for "a man running down the street" the AI must invent the man, the running and the street at once. — 00:00:39 · keyframe
★★ Adil says a bad start frame, blurry or with plastic-looking AI skin (low resolution matters too), carries its flaws through every second of the video. — 00:01:12 · keyframe, AI look

00:01:25 — Mistake 2: a bad start frame, and making one in Soul 2

★★★ Against blurry or plastic-skinned start frames Adil uses the right model and prompts: he selects Soul 2.0 in Images, which he calls hands down the best model right now for production video images. — 00:01:38 · Soul 2.0, keyframe, AI look
★★ For a Soul 2.0 start frame Adil describes the image in simple words ("A man skating") with Higgsfield's prompt enhancer turned on, which expands it. — 00:01:58 · Soul 2.0
★ Adil uses his own Soul ID, trained on over 20 images of himself, for his Soul 2.0 start frame. — 00:02:05 · Soul ID, Soul 2.0
★★★ Adil says the start frame is responsible for 80% of the result: with a bad start frame no amount of prompt engineering saves the video. — 00:02:12 · keyframe

00:02:30 — Mistake 3: vague prompts; who, where, what, camera move and mood in Cinema Studio

★★★ Against vague video prompts, Adil says you need clear intent rather than more words: decide who, where, what is happening, what camera move and what mood. — 00:02:31
★★ In Higgsfield Cinema Studio Adil selects a camera movement instead of prompting for it, for a more accurate result; his favourite is handheld, whose slight shake makes a scene more natural. — 00:03:36 · Cinema Studio
★★ For the mood of a shot, Adil selects one of the genres in Higgsfield Cinema Studio. — 00:03:52 · Cinema Studio

00:04:24 — Mistake 4: character inconsistency; character sheet and tagged character

★★ Adil says inconsistency across stitched clips (an actor's hair or eye colour changing, a birthmark moving) is noticed by viewers even when subtle. — 00:04:27
★★★ To keep a character consistent, Adil makes a character sheet in Nano Banana Pro, uploads it to Higgsfield Cinema Studio with a name and description, and tags the character in each prompt. — 00:04:50 · character sheet, Nano Banana Pro, Cinema Studio
★★ In Adil's Cinema Studio multi-shot with a tagged character, the character showed no face drift or morphing and stayed the same person from different angles and from the back. — 00:05:12 · multi-shot prompt, Cinema Studio

00:05:43 — Mistake 5: overloading the scene

★★ Adil says AI video models replicate real-life physics well but can do only so much at once: too many events in one scene overwhelm both the model and the viewer. — 00:05:43
★★★ Adil advises one primary action per scene plus one or two secondary ones: an explosion as a wide shot, an earthquake as a medium shot, a conversation as a close-up, stitched in an editor or prompted scene by scene in a multi-shot. — 00:06:10 · multi-shot prompt
★★ Adil says some video models are great at physics and others handle human emotions better, and advises testing them all for a given use case. — 00:06:32

00:06:44 — Pro tip: let Gemini or ChatGPT write the animation prompt

★★★ Adil's prompt hack: bring the image to animate to Gemini, ChatGPT or another LLM, describe the animation simply and ask for a detailed prompt, then read and tweak it rather than copy-paste it blindly. — 00:06:52 · Gemini, ChatGPT
★★ Gemini often starts animating the image instead of writing the prompt, so Adil ends his request with "Do not generate, just give me the prompt" in capitals. — 00:07:12 · Gemini

00:07:25 — Pro tips: motion control, detailed image prompts, slow-mo for fast motion

★★★ For pinpoint action or camera movement, Adil records a quick video of himself performing it, transfers it onto the character with motion control and turns on scene control (as heard); it does not solve every problem. — 00:07:25 · motion transfer
★★ When generating images, Adil advises giving the model as much information as possible: clothing, makeup, lighting and composition, as in a real shoot. — 00:07:46
★★ Adil says that to get ahead of most people one must stop thinking like a prompt engineer and think like a movie director. — 00:07:58

“You need to channel your inner movie director.”
★★★ When a super-fast motion keeps morphing, Adil generates a slow-motion version of the scene and speeds it up in any editing tool, which he says saves time on complex action scenes. — 00:08:05

00:08:22 — Recap, giveaway and sign-off

★★ Adil says AI video technology is moving fast but the principles of good cinematography stay the same. — 00:08:30

Threads in this video

restate · Should a video clip start from an image or from a text prompt? — Adil calls typing a long prompt and hitting generate, text-to-video, the biggest mistake: basically a slot machine with zero control over composition; the fix is to use a start frame. (00:00:23)
refine · Should a video clip start from an image or from a text prompt? — Adil says the start frame is responsible for 80% of the result: with a bad start frame no amount of prompt engineering saves the video. (00:02:12)
refine · Which image model should be used for keyframes, edits, text and outfits? — Against blurry or plastic-skinned start frames Adil uses the right model and prompts: he selects Soul 2.0 in Images, which he calls hands down the best model right now for production video images. (00:01:38)
restate · Should prompts be short or long? — Against vague video prompts, Adil says you need clear intent rather than more words: decide who, where, what is happening, what camera move and what mood. (00:02:31)
refine · Should prompts be short or long? — When generating images, Adil advises giving the model as much information as possible: clothing, makeup, lighting and composition, as in a real shoot. (00:07:46)
refine · How do you keep a character the same across shots? — To keep a character consistent, Adil makes a character sheet in Nano Banana Pro, uploads it to Higgsfield Cinema Studio with a name and description, and tags the character in each prompt. (00:04:50)
refine · How do you direct a multi-shot sequence? — Adil advises one primary action per scene plus one or two secondary ones: an explosion as a wide shot, an earthquake as a medium shot, a conversation as a close-up, stitched in an editor or prompted scene by scene in a multi-shot. (00:06:10)
restate · How do you get precise motion (dance, action, choreography)? — For pinpoint action or camera movement, Adil records a quick video of himself performing it, transfers it onto the character with motion control and turns on scene control (as heard); it does not solve every problem. (00:07:25)
first on record · How do you generate fast motion without morphing? — When a super-fast motion keeps morphing, Adil generates a slow-motion version of the scene and speeds it up in any editing tool, which he says saves time on complex action scenes. (00:08:05)
restate · How do you avoid the AI (plastic) look? — Against blurry or plastic-skinned start frames Adil uses the right model and prompts: he selects Soul 2.0 in Images, which he calls hands down the best model right now for production video images. (00:01:38)

Mentioned

Central: Cinema Studio · keyframe · Soul 2.0 · text-to-video

Also: Concepts: AI look · character sheet · motion transfer · multi-shot prompt

Products: ChatGPT · Enhancer · Gemini · Kling Motion Control · Nano Banana Pro · Soul ID

Replied to by

← Previous: The End of Copy-Paste Prompts #aitools #aiart · Next → Higgsfield Audio: A New Feature That Speaks Volumes. The Ultimate AI Text-to-Speech, Voice Swap, and Video Translation Tool