How many reference images can I upload to SpicyFrame?
Up to 12 JPG, PNG or WEBP images under 10 MB each. Most scenes work best with one to four compatible references that each have a clear role.
SPICYFRAME · PRACTICAL GUIDES
Use multiple reference images for AI video: pick compatible photos, give each an @Image role and write one clear prompt. Examples, fixes and FAQs.
5 seconds · 480p · $3.99. Pay per video, no subscription required.
Updated

A public SpicyFrame example illustrating the guide. Results depend on the photo and prompt.
To use multiple reference images for AI video in SpicyFrame, upload one to 12 compatible JPG, PNG or WEBP photos under 10 MB each, give every file a clear job with its @Image number, then write one action, one camera move and one light description. Two to four well-chosen references usually guide a scene better than 12 loosely related ones. Test five seconds in Medium 480p before ordering a longer or High 720p version. References guide the model; they never guarantee a pixel-for-pixel copy of any face or photo.
A reference is most useful when it supplies information the others lack. The lead image usually carries identity: a sharp, front or three-quarter view of the adult subject. A second image can show the complete outfit, a profile angle the planned turn will reveal, or full-body proportions for a walk. Another can describe the setting, a lighting mood or a prop that must appear in the scene. If two images answer the same question, keep the clearer one. If an image answers a question your shot never asks, such as a back view for a front-facing portrait, leave it out.
Use one image for a subtle portrait movement or a slow camera move around an existing composition. Use two when you need identity plus one other element, such as a specific outfit or a room. Use three or four when the movement reveals new angles, for example a turn from front to profile or a walk that shows the full body. Beyond that, each extra file adds more chances of conflict in hairstyle, makeup, age presentation or lighting. SpicyFrame accepts up to 12 references, but that number is a ceiling, not a target. More images do not lengthen the video, add shots or train a saved character between orders.
Lay your candidate photos side by side and remove anything that contradicts the lead image. Keep the same adult, the same period, the same hairstyle and a similar lighting family. Delete near-duplicates, heavily filtered copies, images with text overlays or watermarks and crops that hide the detail you want preserved. Prefer camera-roll originals to screenshots and check that each file is a JPG, PNG or WEBP under 10 MB. Every identifiable real person in the set must be 18 or older and must have agreed to the AI transformation; owning a file does not establish that consent. A curated set of three is faster to diagnose than a folder of twelve.
Start the prompt with the role list, then describe the shot. A reliable structure is: roles, one visible action, one camera instruction, light and mood, and what must stay stable. Two images: “@Image 1 for face and hair, @Image 2 for the outfit. The adult subject turns slowly toward the window, locked medium shot, soft daylight from the left, preserve face and garment.” Three images: “@Image 1 for identity, @Image 2 for the right profile, @Image 3 for the hotel corridor. The subject takes two slow steps and looks over the shoulder, gentle side tracking shot, warm practical lights, quiet footsteps.” Name each role once and keep the sentences short.
A four-image set suits an outfit reveal where face, garment, setting and mood all matter. Prompt: “@Image 1 for facial identity, @Image 2 for the full outfit and shoes, @Image 3 for the rooftop location, @Image 4 for the golden-hour light and colour. The adult model takes one confident step and settles into a pose, slow low-angle push-in, warm sun behind the subject, light breeze in the fabric, distant city ambience, preserve face, outfit and proportions.” Check that the rooftop and the outfit photo were shot in compatible light. If @Image 4 is a mood image of a different person, say it is for colour only so the model is less likely to borrow that face.
Before: “Use all my photos and make her walk somewhere nice, cinematic, beautiful, keep everything the same.” This gives no roles, no camera and an impossible instruction. After: “@Image 1 for face, @Image 2 for the black satin dress, @Image 3 for the marble lobby. The adult subject walks two slow steps toward the camera and pauses, steady backward tracking shot, warm evening light, soft footsteps, preserve face and dress.” The rewrite says what each file contributes, what happens and how the camera behaves. If a fourth photo adds nothing to that shot, leaving it out is an improvement, not a loss.
The @Image labels follow the order of the references in the studio. Before submitting, compare each numbered thumbnail with the role you wrote: a prompt that calls the dress @Image 2 when the dress is the third file will send the wrong instruction. Re-read the numbers after adding or removing a reference. Keep the lead identity image first, and when you reuse a set for several clips, paste the same role line unchanged so outputs stay comparable.
Run the first test at five seconds in Medium 480p, which costs 399 piments ($3.99). A five-second High 720p version costs 799 piments, and Ultra High 1080p costs 1,999. Review the clip, decide which element failed and change only that element in the next attempt: remove a conflicting reference, simplify the action or calm the camera. Changing the set, prompt and camera together hides the cause. Each new render is a separate paid request with its own price shown before checkout. When the five-second test passes, order the chosen duration and quality with the same references and the same role line.
If the face looks like a blend of two people, remove older photos, retouched versions or mood images of other people. If the outfit switches during the clip, keep a single outfit reference and name it explicitly. If the setting ignores @Image 2, the location may contradict the lead composition; choose a setting photo with similar camera height and light direction, or move the scene into the room already shown in @Image 1. If colours shift, two references probably disagree about the light. If hands distort, the issue is usually motion, not the number of images; reduce the gesture.
For a blink, a small smile, breathing, hair moving in a breeze or a slow push-in on an existing composition, a single clean photo often conveys the whole scene. Add another image only when it supplies information the first lacks, such as a hidden profile or the full outfit. You can also generate from a prompt alone, although your references are what keep the face, outfit and setting anchored. Review faces, hands and clothing in the completed video before downloading or sharing it.
| Reference role | What the photo should show | Prompt label example |
|---|---|---|
| Identity (lead) | Sharp front or three-quarter face, even light, no sunglasses | @Image 1 for face and hair |
| Profile angle | The same adult from the side the turn will reveal | @Image 2 for the right profile |
| Full outfit | Complete garment and shoes in clear, uncropped light | @Image 2 for the exact outfit |
| Full-body proportions | Head to feet with space around the limbs | @Image 3 for body proportions |
| Setting | Location shot from a similar camera height and light direction | @Image 3 for the hotel corridor |
| Light or colour mood | A frame whose light and palette you want to borrow | @Image 4 for colour and light only |
Up to 12 JPG, PNG or WEBP images under 10 MB each. Most scenes work best with one to four compatible references that each have a clear role.
Not automatically. A compatible profile or three-quarter view can help a turn, but mixed ages, hairstyles or retouched versions can blend into a less accurate face.
Use the @Image numbers shown in the studio, which follow the upload order, and give each one a role, such as “@Image 2 for the outfit.”
You can describe it as a colour and light reference only, but keep it out when possible, because the model may borrow facial traits from it.
No. Each SpicyFrame generation is one continuous vertical scene. Plan several clips and edit them together if you need a sequence.
Signed-in users can keep saved references when storage is enabled, but the model does not learn a permanent character. Reuse the same set and role line for each new clip.
Only images you own or may use. Every identifiable real person must be 18 or older and must have agreed to the AI transformation.