AI characters and prompts
Quick answers on keeping one AI character recognisable across scenes and on writing prompts that give a video model less to guess. Covers reference images in Nano Banana Pro, the Seedance reference system, Elements, structured and JSON prompts, and debugging a bad generation. Facts come from Google, Higgsfield and LTX pages checked on 2026-10-22.
How do I make a consistent AI character?
Design one character image first, then reuse it as a reference in every later generation. Google's docs say gemini-3-pro-image supports up to 5 character images with high fidelity and up to 14 reference images in total. Describe hair, clothing and distinguishing features the same way every time.
Read moreCan Nano Banana Pro keep the same character across images?
It supports reference-image workflows built for this. Google documents up to 5 character images at high fidelity within up to 14 references for gemini-3-pro-image. References improve consistency but do not guarantee it, so compare every output with your original design before you animate it.
Read moreWhat is a character reference sheet?
A reference sheet shows one character from several angles so later prompts have something to match. Google's docs show building a 360-degree view by prompting for one angle at a time and feeding earlier images into later prompts, and say complex poses need a reference image of the pose.
Read moreWhat is the difference between Nano Banana Pro and Nano Banana 2.1?
Google's docs call Nano Banana Pro the premium choice for the most complex visual tasks, with the highest world knowledge and precise control. Nano Banana 2.1 is the high-efficiency workhorse with Flash-level speed and cost, and the docs recommend it for all new projects of that kind.
Read moreHow many reference files can Seedance take?
Higgsfield says Seedance 2.0 combines up to 9 images, 3 video clips and 3 audio files in one generation. Seedance 2.5 accepts up to 50 references: 30 images, 10 video clips and 10 audio files. Limits are per generation and can change, so check the help page.
Read moreHow do I keep one character across separate Seedance generations?
Save the character as an Element. Higgsfield says references hold a face within a single generation, but for the same character across separate generations you click Elements under the prompt field, or type @, and select the character, with no re-uploading for each clip.
Read moreWhat do @character, @style, @motion and @audio do?
Higgsfield's reference system: @character holds face geometry, skin tone and style consistent within the clip, @style applies lighting, palette and mood from an image, @motion copies camera behaviour from a video clip, and @audio syncs visuals to audio, including lip sync and ambient sound.
Read moreWhy does my AI character look different halfway through a clip?
Higgsfield suggests adding more specific visual anchors such as hair colour, clothing details and distinguishing features. For consistency across generations, reference the character as an Element. Change one thing per run so you can tell which edit fixed it.
Read moreWhat is a JSON prompt for AI video?
A JSON prompt organises instructions into named fields such as camera, subject, scene and timing instead of one paragraph. LTX describes it as giving the model a production brief rather than a casual description, aimed at more predictable results for complex scenes.
Read moreIs a JSON prompt better than a plain text prompt?
Only for complex scenes, according to LTX, which says plain language works well for simple single-element generations and becomes ambiguous with several elements, spatial relationships or timing. LTX Studio parses JSON, but support varies by model, so test it on yours before relying on it.
Read moreCan I use structured prompts in Seedance?
Higgsfield says Seedance 2.5 follows prompts noticeably better than 2.0, so structured prompts translate into the shot more reliably. Its recommended order is subject and action, setting and lighting, camera move, then mood or style. The article we read does not require a JSON format.
Read moreWhat order should I write a video prompt in?
Higgsfield's structure for Seedance is subject and action first, setting and lighting second, camera move third and mood or style last. Keep the first generation short, because image-to-video is more predictable than text-only and uses fewer credits while you learn how the model responds.
Read moreWhich camera terms does Seedance understand?
Higgsfield lists terms Seedance reads directly: dolly in, truck left, arc shot, push in, pull back wide, handheld follow, crane up and orbital move. Its advice is to replace vague wording with specific terms when the output ignores the camera instruction.
Read moreWhat reference image works best for the first generation?
Higgsfield recommends a sharp, front-facing, well-lit reference image with a short prompt for the action and camera move, run as image-to-video at 720p. Most quality problems show at the 5 to 8 second mark, so check whether motion holds throughout and the camera move happened.
Read moreWhat does Seedance 2.5 Edit do?
Higgsfield describes it as a separate model that changes a clip you have already generated, including region-level edits where you mark the area to change on the frame, and audio. It runs at 480p to 720p and matches the source clip's length.
Read moreCan I use a real person as an AI character?
Higgsfield says to train a Soul ID first for a real person and create the Element from its portraits. Consent matters: YouTube requires disclosure when content makes a real person appear to say or do something they did not, or generates a realistic scene that did not occur.
Read moreDo I need to train a model to keep a character consistent?
Not necessarily. Google's docs describe reference-image workflows, and Higgsfield's Elements store a character for reuse without retraining. A Soul ID is Higgsfield's route for real people. Training custom models is another route some creators use, but we have not tested it.
Read moreCan Claude Code help write AI video prompts?
Yes, as text: it can draft prompts in a fixed structure and check them for gaps before you spend credits. Claude's pricing page lists Claude Code on Pro and Max, not Free. It writes the prompts but does not generate the video.
Read moreHow do I debug a bad AI video generation?
Change one thing per run. Higgsfield maps symptoms to fixes: vague camera wording gets specific terms, character drift gets visual anchors or an Element, and a jerky cut gets first and last frame references. Prototype at 720p so each test costs fewer credits.
Read moreWhere can I get copy-ready prompt templates for story videos?
AI Video Empire includes copy-ready templates in every lesson: a cast lock, scene JSON, a script block, ending clips, image prompts for a four-angle photo shoot, and a prompt checker. It is a one-time payment for lifetime access, and the AI tools have their own plans.
Read more