PTM Prepa-Tec Media

Week 2, day 4: animating stills

Making a still picture move.

Students learnt the basics of cinematography first, meaning camera angles, camera movements and shot types, then how to effectively transform their own images into animation: describing subject action, camera motion and background movement as three separate instructions, and judging each result against the still it came from. They also learnt skills like moodboards and LoRAs, small training files that keep a character or a style the same, to work on AI video’s biggest issue, consistency between frames. The shots are shown large, with the frame they started from beside them.

The lab book Week 2, day 4 · skills Before & after Week 2, day 5 · re-rendered shots
Explore
Content generated from AI through an API connection to today's class' Slack channel.

Week 2 · Day 4 · #w2d4_animating-stills

The class lab book

We split into teams, each hacking a different corner of an AI animation tool, then reported back what actually worked. This is what the class now knows, and how we found out.

Pick a skill to test Experiment (teams try different approaches) Report the result Learn: feed it to the trailer

Experiment 01· A/B · 2 teams

Camera Control

Can you direct a camera move just by asking for it in the prompt?

Worked

▸ Prompt: "do a zoom out"

It was a little bit difficult but we told to the ai to do a zoom out.

Mariana T.· You Already Knew

Worked

▸ Zoom + a tighter prompt

in this video we learned how to use the zoom of the camara an how to make god prompts

Indira· Nopal in Seol

What the class learned

Yes. Name the move ("zoom out") and it obeys. Phrasing takes a couple of tries, so pair the move with an otherwise clean prompt.

Experiment 02· A/B · 2 teams

Prompt Economy

Does packing more into one prompt give you more control?

Didn't work

▸ Many concepts, one prompt

Terrible, We learned that mixing a lot of concepts in one simple prompt can confuse the AI, thus creating weird stuff

Noé M.· The White Hanbok

Didn't work

▸ Loose, unfocused prompt

we had problem here because of a bad prompt we did, and started making so many strange faces

Connie C.· Hollowskin Group

What the class learned

No, the opposite. Two teams overloaded their prompts and both degraded (weird artefacts, distorted faces). Say one thing clearly; add detail only once the base shot holds.

Experiment 03· A/B · 2 teams

Fixing a Detail Without Re-rolling

When one detail is wrong, must you regenerate the whole shot?

Fixing a Detail Without Re-rolling Worked

▸ Edit tool on the wrong region

We had difficulties with the nationality of the characters because they are suppose to be korean but the AI sometimes makes them look like latinos so we used the edit tool to make them Korean

Aylin R.· BTS Project

Fixing a Detail Without Re-rolling Worked

▸ Select the whole area + retry

Try selecting everything, not just one part, so the AI doesn’t only work in one area. Also, if it doesn’t work the first time, try adding a little more information and do it again.

Aaron· Dreemteem

What the class learned

No. Mask the region and regenerate only it. The compounding tip from the second team: select the whole area (not a sliver), and add detail on a retry if the first pass misses.

Experiment 04

Start + End Frame

How do you get controlled motion instead of a random animation?

Worked

▸ Two stills as start & end frame

we created this scene with 2 images creating them with the start and end frame

A student

What the class learned

Give the model an explicit first and last frame, and it interpolates between them. More predictable than animating a single still.

Experiment 05

Motion Transfer

Can you map one clip's movement onto another?

Didn't work

▸ Transfer motion between clips

Terrible Motion Transfer

Ian F.· Echo

What the class learned

Not reliably in Krea right now. The result was unusable. Animate a fresh generation instead, and re-test this feature next month.

Experiment 06

Character Realism

How close can the AI get to a believable human performance?

Mixed

▸ Push for a humanlike character

We are able to make characters look more humanly, but we need to act and record ir better

Mila D.· The Last Station the Cure

What the class learned

It reaches humanlike faces, but believability tracks the reference, meaning the performance and recording you feed it. Direct and record the source better.

Experiment 07

Training the Model

Can you teach the model your own subject before generating?

💬 1 Still testing

▸ Trial the training feature

Here we were playing and trying the function of training the AI

Connie C.· Hollowskin Group

What the class learned

Team started testing custom training, with no verdict yet. Flagged for a proper A/B next session.

Experiment 08

Stretching Credits

How do you make paid video credits last?

Stretching Credits Worked

▸ Free tool for stills, paid for motion

using chatgpt to generate images so i could use more credits for video generation

Mateo M.· El Tigre de Taekwondo

What the class learned

Generate the stills in a free tool (ChatGPT) and spend paid Krea credits only on the video step. A workflow hack the whole class can reuse.

Experiment 09

Tool choice: lighting

Is every tool good at everything?

Tool choice: lighting Didn't work

▸ Generate a lit frame in Grok

Horrible shadow - Grok

Ivi· Hollowskin Group

What the class learned

No. Grok produced bad shadows. Route lighting-critical frames to a different tool.