How to create a short AI film that looks like a spontaneous recording, and then change the outfit, object, or location without rebuilding the shot from scratch? This guide walks you through three stages: the opening image in Higgsfield Soul 2.0, animation with references in Seedance 2.5, and corrections of the finished clip in Higgsfield Genjutsu. You get a work plan, original prompts to copy, and a checklist of things to verify before exporting.
Tool status verified on October 4, 2026. Interface, limits, and model availability may change. Prompts are examples for adaptation, and generators do not guarantee exact execution of every instruction.
The Entire Process at a Glance
| Stage | Tool | Result |
|---|---|---|
| 1. Appearance | Soul 2.0 | An image that establishes the composition, lighting, texture, and color of the first frame |
| 2. Motion | Seedance 2.5 | A film with this image as the starting point, featuring camera and character movement |
| 3. Change | Genjutsu | A variant of the existing film with the selected element swapped |
In this example, we create a seven-second shot: a person exits a small café onto the street, the camera follows her, and after exporting, we change the color of her jacket. This is a custom practice scene. You can substitute your own location, character, and product.
Prepare Materials Before You Start
- Define the Format and Purpose
For example, a vertical 9:16 film for social media, seven seconds long, one shot without cuts. The initial image and resulting film should have matching proportions.
- Gather References
Prepare a separate photo for the first frame, photos of the character's appearance, and a photo of the outfit. Give them simple names:
start.jpg,face-front.jpg,face-profile.jpg,jacket.jpg. Use materials for which you have rights and permissions from the people visible in the frame. - Record What Must Remain Constant
In this exercise: the same person, door layout, clip length, one crossing of the threshold, and no additional characters. This makes it easier to assess the result of each generation.
Stage 1: Build the First Frame in Soul 2.0
Enter Higgsfield: Image → Soul 2.0. The tool's documentation confirms that you can combine the prompt with settings for style, character, and colors. There is an important detail here: after adding a reference image, the prompt field may become unavailable. If you want to use the text below, start generating without a reference image; if you are working from a sample photo, guide the appearance through that image and the available settings.
A good prompt for a photo describes a specific type of camera, light source and its effect, and image imperfections. Simply saying "realistic, cinematic" gives the model fewer clues than describing reflections, noise, and shadows.
A candid photo at the doorway of a small neighborhood café on an overcast afternoon. One adult woman in a faded green jacket stands half a step inside, turned toward the sidewalk, one hand near the door handle. Wet pavement, a bicycle and soft street reflections outside. Framed at chest height, with room for her to walk through the doorway. Early-2000s compact digital camera look: direct on-camera fill light brightens her face and the painted doorframe; the café interior remains darker, while the street beyond is slightly hazy. Visible natural skin texture, modest sensor noise in the shadows, small hot highlights on wet surfaces, mild bloom around the doorway, imperfect candid composition. Vertical 9:16.
Generate several variants and choose the one where passing through the door is physically possible: the entire silhouette needed for movement is visible, the door handle is in the right place, and the sidewalk connects seamlessly with the interior. Looking good in a still image is not enough; the frame must be the starting point of movement.
Stage 2: Bring the Image to Life in Seedance 2.5
Upload the selected image as the first frame, then add the remaining references. Seedance 2.5 supports generation with multiple references and sound; the exact way to label files depends on the interface you are using the model in. Check the order of attachments before pasting the prompt. If the interface uses a different syntax than @Image 1, replace the labels with its markings.
| Attachment | Role | What Not to Transfer |
|---|---|---|
| @Image 1 | First frame, café layout, lighting | None |
| @Image 2 | Identity of the character's face | Background of the portrait photo |
| @Image 3 | Profile and shape of the hairstyle | Clothing from the profile photo |
| @Image 4 | Cut and color of the jacket | Model and environment from the clothing photo |
First, tell the model what each file is for. Then describe the motion in chronological order. Finally, add a few critical rules. There is no need to copy dozens of pages of instructions: with a short clip, it is easier to correct one visible problem at a time.
REFERENCE MAP
@Image 1: exact opening frame, café doorway, street, framing and light.
@Image 2: the woman's face and identity only.
@Image 3: the same woman's profile and hairstyle only.
@Image 4: her faded green jacket only; do not copy the person or background in this image.
Create a 7-second vertical video in one continuous handheld take. At 0.0 seconds the frame matches @Image 1. The woman starts walking through the café doorway. The camera follows from a step behind at chest height, with small natural drift and a slightly late correction. From 0 to 3 seconds she reaches the door and opens it; from 3 to 5 seconds she crosses the threshold; from 5 to 7 seconds she continues onto the wet sidewalk. Her hair moves more in the outdoor breeze. The phone's automatic exposure gradually adapts to the brighter street; the light sources themselves remain constant. Natural sound only: footsteps, a soft door sound, distant street noise and a little wind on the microphone.
Keep her face, jacket and body consistent. Preserve visible skin texture and modest sensor grain. Every movement has a visible start, middle and end. No cuts, teleporting, digital zoom, slow motion, music, captions, extra people or sudden changes in lighting. End while she is still walking.
Why This Works: “exact opening frame” anchors the beginning, time intervals organize the action, and the final rules protect against common errors: position jumps, face changes, and unwanted cuts. Camera movement, focus, exposure, and sound are described as camera behavior, not as editing effects.
Post-Generation Check
- Compare frame 0 with the source image: the position of the person, the door, and the light should match.
- Watch the crossing of the threshold frame by frame. Feet, hand, and door should not “jump” between positions.
- Check the face and clothing at the beginning, at the threshold, and at the end.
- Listen to the sound: do the footsteps and door match the visible movement, and does no music or extra voice appear?
- If one thing failed, only correct the relevant sentence in the prompt and regenerate. A long set of conflicting prohibitions makes diagnosis difficult.
Stage 3: Change the Finished Film in Genjutsu
Genjutsu works on an existing clip. The official description distinguishes between Object Swap, which changes the specified element, and Motion Transfer, where the motion from the film is used to create a new version with a different character or scene. This is the tool's purpose, not a promise that every detail of the original will be perfectly preserved. The result must be watched throughout the duration of the film.
Mode A: Swapping One Element
Upload the finished film as @Video 1 and the photo of the new jacket as @Image 1. Indicate the person by their position in the frame, name the change, and specify what should remain. A short prompt is easier to evaluate than a re-description of the entire scene.
Replace the green jacket worn by the woman walking out of the café in @Video 1 with the navy jacket from @Image 1. Keep her face, movement, camera motion, doorway, street, timing and sound the same.
If you also want to change the café sign, create a second variant and add a separate sentence for the sign. With each additional change, the risk increases that the model will also modify adjacent elements.
Mode B: Motion Transfer
When rhythm and action layout are most important, use the film as a motion template and the photos as references for the new character and location. Unlike the jacket swap, a far-reaching new version of the image may emerge. When using someone else's film, ensure you have the right to use it as source material.
Use the walking action, camera path and timing from @Video 1. Recreate the woman using @Image 1 and the location using @Image 2. She exits the doorway and continues walking in the same direction. Keep the shot continuous and preserve the original pace.
Do not assume that the label @Image 1 always refers to the same file. The numbering pertains to the current task and order of attachments. Check it again before running.
Common Problems and Quick Fixes
| Problem | What to Change |
|---|---|
| The film does not start with the selected image | Repeat in the first sentence that 0.0 s should correspond to @Image 1; remove conflicting descriptions of the new frame. |
| The face or outfit changes during the film | Separate identity references from wardrobe references and name the role of each file. |
| The person suddenly appears behind the door | Break down the transition: approach, contact with the door, crossing the threshold, further step. |
| Exiting looks like a dimming of light | Describe the gradual adaptation of automatic exposure and constant light sources. |
| Genjutsu also changes the background or face | Specify one element by its position in the frame and explicitly list the features that should remain. |
| A fragment of dialogue is missing or the sound does not match | Shorten the speech, remove excess simultaneous actions, and generate a shorter, simpler variant. |
Checklist Before Publication
- Check the first and last frame and the moment of greatest movement.
- Watch the entire clip without sound, then listen to it without looking at the image.
- Compare the version before and after Genjutsu: did only the planned element change?
- Verify rights to photos, film, voice, trademarks, and character images.
- Keep the source file, prompts, order of references, and finished variants. This will facilitate later corrections.
The simplest start is one person, one shot, one movement, and one later change. Once this variant works, add more people, dialogue, props, or locations one at a time.
Comments (0)
No comments yet. Be the first!