From Idea to Image, Part Eight

Prepared with Claude; edited and published by Owrang.

About this series

From Idea to Image is a weekly self-study course in making images with AI. The text of the series is also written with the help of AI and is edited and published by Owrang.

Each part teaches one skill and ends with a simple exercise you can do the same day. From this part on, each time we teach one task with its real tools. If several models can do that task, we name them all and compare them.

By now we know Mahin, a sixty-year-old woman with a green cardigan, round glasses and a white streak in her hair. We have sat her at the kitchen table, and in parts six and seven we looked at the principles of movement. Now we want Mahin to turn her head and look at the window. It is a simple movement that is not so simple. The face must stay the same face, the glasses must not shift, the window must stay where it is, and the head movement must look natural.

Which tool should we choose to make these five seconds?

This week we weigh five video models against the needs of a single shot. The aim is not to know everything they can do. We want to learn how to tell, before we start, which tool will get closest to the result we want.

This week's lesson

Choose the model by the shot, not by its fame. Before you open any tool, know what your shot needs. Do you only have a starting image, or must the shot also end in a particular position? How many seconds is it? Must the character stay the same across several shots? Is the body movement large, or only the head and the gaze? Does it need sound too? The answers to these questions choose the model.

Before making the video, a shot card

For each shot, write on a small card:

First frame: which image.
End position, if it matters.
Length: how many seconds.
Camera: still or moving, and what movement.
Character movement, in one sentence.
What must stay the same, for example the white streak of hair, the glasses, where the window is.
Where the characters and objects are in the frame, and which side of the imaginary line they are on.
The direction of the gaze, the direction of movement, and the point the gaze must reach.
Sound: yes or no.

This card is what becomes the prompt, and afterwards you judge every clip against it.

Five models for bringing an image to life

What follows comes from the makers' own official pages, which we read on 8 November 2026. These are the features the makers announce. Our own experience with some of these models is in the section "What happened in practice", and the two are not always the same.

One more important point: a model does not have the same features everywhere. The same model on its maker's site, in Runway or in the programming version (API), and on different plans, may offer different lengths, quality and inputs. Runway itself offers several other companies' models, and sometimes has an older and a newer version of the same model side by side. So before you start, check the guide of the place where you are actually working. Tools also change quickly. The cost of each second also varies a great deal from model to model, and from one quality setting to another. Before the final version, make a short version or a low-quality preview and look at the result.

Runway Gen-4.5. It takes an image as the first frame and a prompt, and makes a clip of 2 to 10 seconds, at 720p and 24 or 25 frames per second. According to Runway's guide, for now it accepts only text and images, and other inputs are "coming soon". It has no last frame and no reference image. Camera movement can only be described in the prompt, and Runway says the model follows complex, sequential instructions well. The guide does not mention sound. Access and practice: the Runway site.

Kling 4.0. Available to everyone since 15 October 2026. It accepts a first and last frame, up to 10 keyframe images through the clip, up to 10 reference images, up to 5 reference videos and up to 7 "elements" for keeping characters and objects consistent, but at most 15 references in total. Clips run from 3 to 30 seconds, at 720p, 1080p or 4K. It also makes two-channel sound and dialogue, but Persian is not in its list of languages. Kling reports better realism, stable large movements and better lip sync. The lighter version, Kling 4.0 Flash, goes up to 20 seconds at 720p. Access and practice: the Kling site. Runway's API guide also lists Kling 4.0 and Kling 4.0 Flash among its image-to-video models.

Gemini Omni 1.2 Flash. Google introduced it on 5 November 2026. It takes text, images and video, up to 6 reference images, that is up to three images per character and up to two characters, and a first and last frame. Clips run from 3 to 15 seconds and can be extended to 30 seconds. The default output is 720p, with 1080p and 4K by upscaling. It makes sound together with the picture, but according to its guide only English is fully supported. An important point: by default it makes a video with several different shots, and if you want one continuous shot you have to say so explicitly. Google says the model is still improving at fast sports movements and complex physics. All outputs carry the invisible SynthID watermark. Access and practice: Google AI Studio. Omni Flash is also in Runway, but with different features. According to Runway's guide, clips there run from 3 to 10 seconds, always at 720p, and it takes a first frame and up to 10 reference images. The guide does not mention a last frame. Runway's API guide lists older versions and version 1.2 side by side. Check which version you are using, because their features are not the same.

Seedance 2.5, made by ByteDance. In a single generation it accepts up to 30 images, 10 videos and 10 audio files as references, and makes a clip of up to 30 seconds, with several continuous shots and sound, which can be extended. The maker sees its strength in long storytelling and in keeping the faces and voices of several characters, and says itself that complex movements and crowded scenes are still difficult. It has not announced which languages its sound supports. Access and practice: Runway. According to Runway's guide, clips there run from 4 to 30 seconds, at 480p, 720p or 1080p, and sound can be turned on or off. Its Keyframe mode accepts a first frame and a last frame.

Luma Ray3.2. It accepts a first and last frame and up to 16 keyframe images in one clip. Clips are 5 or 10 seconds, up to 1080p, with 16-bit HDR output for professional colour work. Camera movement can be taken from a reference video. It makes no sound. The Ray3 guide also says character consistency "may vary". Access and practice: the Luma site.

There is one more way: movement from your own performance. Runway Act-Two takes an image of the character, together with a video of you or whoever plays the role, from 3 to 30 seconds, and gives that person's face, hand and body movement to the character. The output is 720p at 24 frames. Facial expressiveness runs from 1 to 5, with 3 as the default. The hand and body movement option only works when you give the character as an image. The guide says to frame the performer at least from the waist up, without cuts, with one person in the frame. Act-Two has no control at all over the direction of the gaze.

Which model for which shot

A short shot with one simple movement from one image, such as a blink or a turn of the head: Gen-4.5 is the simplest way. Omni is good too, provided you write "one continuous shot, no cuts".

A shot that must end in a particular position so that it joins the next shot: a model that accepts a last frame, that is Kling 4.0, Luma, Seedance 2.5 in Runway, or Omni in Google AI Studio.

A character who must stay the same across several shots: the models that take reference images and "elements", that is Kling 4.0, Omni and Seedance. Give them the character sheet from Part Seven.

A long shot, from 15 to 30 seconds, or several continuous shots in one generation: Kling 4.0 and Seedance 2.5.

Natural body and hand movement, such as pointing or giving a speech: Act-Two with your own performance, in a medium shot.

A shot whose colour and light must be worked on in professional editing: Luma Ray3.2 with HDR output.

Dialogue in Persian: Persian is not in the announced language lists of these models. That does not mean none of them can produce Persian, but no reliable support has been announced either. So for Persian dialogue it is better to record the voice separately and test lip sync with the speech separately. Next week we look at exactly that.

Combining models

The best result usually does not come from one model alone. Here is an example of a chain that worked in our own work.

1. Make the image and the character sheet with an image tool, as in the earlier parts.

2. Make the simple shots without speech with an image-to-video model, for example Gen-4.5 for a turn of the head.

3. If a shot must end exactly in a particular position and your model does not take a last frame, make the clip starting from that end position and play it in reverse. This only works for movements that also look natural in reverse, such as turning the head. Blinking, hair, clothes and body movement sometimes look unnatural in reverse, so watch the reversed clip carefully. At the join, try a straight cut and a very short dissolve, and keep whichever looks more natural. Or make the same shot with a model that takes a last frame.

4. For natural body movement, act the role yourself on your phone and give it to the character with Act-Two. The direction of the gaze has to be acted. Next to the camera, at the point where the character should look, put a mark with tape.

5. Make the last frame of one model's clip the first frame of another model's clip. This way two shots, even from two models, meet at the cut. It makes the visual join easier, but it does not guarantee that movement, light, camera angle and the direction of the gaze stay the same.

6. If a 720p shot looks softer next to the other shots, try upscaling it to 1080p with a video upscaling tool. Upscaling does not create real new detail, so decide shot by shot.

7. Lay everything out in the editor and put the room sound under it without a break, the lesson of Part Seven.

Practical exercise, one shot three ways

Step 1. The shot card. For the shot of Mahin turning her head towards the window, fill in the shot card. First frame: Mahin's reference image. End position: profile, facing the window. Length: five seconds. Camera: still. What stays the same: the white streak of hair, the round glasses, the green cardigan, the window on the right of the frame.

Step 2. Gen-4.5. Give Mahin's image as the first frame and write: "Static camera on a tripod. The woman sits silently at the table, lips closed. She slowly turns her head towards the window on the right of the frame and blinks naturally. Her hands rest on the table around the tea glass, still." Set the length to five seconds.

Step 3. The same shot with a last frame. If you have access to Kling 4.0, Luma, Seedance 2.5 in Runway or Omni in Google AI Studio, make the same shot, this time with two images: Mahin's reference image as the first frame, and an image of her in profile facing the window as the last frame. Make the profile image with your image tool and the character sheet. In Omni, remember to write "one continuous shot, no cuts".

Step 4. Comparison. Watch the two clips side by side and write notes for each. Did the face stay the same? Did the white streak of hair and the glasses stay in place? Did the window stay on the same side? Was the head movement natural? Was anything or anyone added to the frame? Was any sound made that you did not ask for?

Step 5. Movement from your own performance. If you have access to Act-Two, film yourself on your phone sitting at a table and turning your head towards the window, from the waist up, five to ten seconds, with bright light on your face. Put a mark next to the window so that your gaze goes exactly there. Give Mahin's image as the character, turn on hand and body movement and set facial expressiveness to 2. Compare the result with the other two clips.

Step 6. Assembly. Put the best clip into last week's scene in place of the look shot, and watch the scene from the beginning.

When the result is not what you expected

Someone or something appeared in the corner of the frame? Before a clip goes into the edit, check it frame by frame, and pay attention to the edges of the frame and to the sound. Some models add people, and models that make sound may make sound you did not ask for.

Omni made several shots when you wanted one? Add "one continuous shot, no cuts" to the prompt.

The face slowly changed during the clip? Make the clip shorter, or use a model that takes reference images and give it the character sheet.

The hands went wrong? Keep the hands out of the frame or resting quietly on the table. For hand movement, Act-Two with your own performance in a medium shot works better than a prompt.

The shot does not match the next shot? Make the last frame of one the first frame of the other, or use a model with a last frame.

You made a shot several times and it did not come right? Do not make it again. Change the model, make the shot smaller, or do that part with a cut.

What happened in practice

Owrang works on visual projects with AI and has used these same models in them. What we learned, with the tools named. These are the results of our tests on particular shots, not settings that always give the best result.

First, Act-Two ruined the hand in a wide shot and worked properly in a medium shot. In the wide shot the character raises a hand, and in the output the hand lost its shape. We made the same performance again in a medium shot, with the character image from the waist up, hand and body movement turned on and facial expressiveness at 2. The hand became natural, and when it rose it went out of the top edge of the frame, just as with a real camera. Level 3 had made the facial expression exaggerated.

Second, repairing a hand in editing. In our experience, several attempts to fix a misshapen hand with Aleph, Runway's video editing tool, gave no usable result. We learned that remaking a shot at the generation stage is sometimes more reliable than trying to repair a complex movement error in editing.

Third, a performer who does not match the character. For a character with a long beard, we gave Act-Two the performance of someone with a short beard. When we gave the character as an image, the face changed and looked into the lens, while the character was in three-quarter view. When we gave it as a video, the beard seemed to come away from the face. Choose a performer who looks as much like the character as possible, and match the angle of their head to the character's angle.

Fourth, Gen-4.5 really did turn the head, but it also made someone we had not asked for. From the image of one shot we made a head turn from three-quarter view to profile, and the face stayed the same. But at the end of the clip a woman appeared at the edge of the frame, standing opposite the character, her lips moving as if she were talking. We kept the movement and set that clip aside.

Fifth, reverse to reach the end position. Gen-4.5 does not take a last frame. For a shot that had to end exactly in the position where the dialogue begins, we made the clip starting from that position and played it in reverse, with a short dissolve at the join.

Sixth, the gaze has to be acted. None of Runway's tools has control over the direction of the gaze. When we wanted the character to look at the crowd, the performer had to act the direction of the gaze by looking at a mark next to the camera.

Seventh, a mechanical face and gaze. In the shots we had made with Seedance, the overall movement was good, but to us the face and the gaze read as mechanical. That is why we made the close-ups of the face in other ways.

Working terms

image-to-video / I2V

text-to-video / T2V

first frame / last frame

keyframes

reference image / element

multi-shot (several shots in one generation)

extend (continuing a clip)

performance capture / motion transfer

facial expressiveness

upscale

invisible watermark / SynthID

This week's exercise

Choose a shot of your own character that has a simple movement. Write the shot card. Make the same shot with two models, one without a last frame and one with a last frame. If you can, make a third version with your own performance. Put the three clips side by side and for each one write what stayed the same and what changed, how long it took you to reach the result, and which one you will put into your scene.

Test yourself

Show the clips to someone without saying which was made with which model. Ask which was more natural, and why. Ask in which one the character stayed the same person as in the image. If their reasons match your notes, you have learned to choose the model by the shot.

Next week. We plan to make the character speak. Tools that sync the lips to your own voice, from Runway and Kling to sync. and LatentSync, and tools that make a Persian voice, will be compared. We will also give the method that worked with Kling 4.0 for matching the lips to a reference video.

Where to practise, and each model's official guide

All these links go to the makers' own official pages. All are in English.

Runway Gen-4.5, guide to creating. https://help.runwayml.com/hc/en-us/articles/46974685288467-Creating-with-Gen-4-5

Runway Act-Two, guide to performance capture. https://help.runwayml.com/hc/en-us/articles/42311337895827-Performance-Capture-with-Act-Two

Kling 4.0, release note. https://kling.ai/release-note/release-notes/Kling_4?type=dialog

Gemini Omni, Google AI Studio guide. https://aistudio.google.com/docs/omni

Seedance 2.5, ByteDance Seed page. https://seed.bytedance.com/en/seedance2_5

Seedance 2.5 in Runway, guide. https://help.runwayml.com/hc/en-us/articles/53542207042323-Creating-with-Seedance-2-5

Gemini Omni Flash in Runway, guide. https://help.runwayml.com/hc/en-us/articles/53031657806611-Creating-with-Gemini-Omni-Flash

Luma Ray3.2. https://lumalabs.ai/ray