AI characters explained

Definition

AI characters are persistent visual identities created by AI models that reuse the same face, body proportions and clothing details in every new generation. They differ from single-prompt outputs because the system stores reference data instead of starting from scratch each time.

How the system keeps identity

Flixly stores a character embedding from the first image you upload. Later generations pull that embedding into models such as Seedance 2.0 or Veo 3.1. The embedding sits between the text prompt and the diffusion process so the model cannot drift far from the original features. When you switch to Kling 3.0 the same embedding file travels with the job.

Seedance 2.0 adds temporal attention layers that check every frame against the stored reference. This step reduces face swaps between shots that last longer than eight seconds. Wan 2.7 uses a lighter version of the same technique for shorter clips under four seconds.

Concrete inputs you provide

You start with one clear reference photo at 1024 by 1024 pixels. Add a short text description that lists age, hair color and clothing. Optional inputs include a 3-second motion clip for body movement or a second angle photo for better depth.

Output files come back as 1080p MP4 at 24 frames per second or static PNG images at the same resolution. Each generation consumes between 8 and 22 credits depending on length and model choice.

Typical workflows

Creators build a character once then reuse it in a full scene sequence. One workflow starts at the Reference to Video page, uploads the reference, then feeds the result into Image to Video for motion. A second path uses the same reference inside Manga Creator to keep the same face across ten panels.

Series producers often chain Anime / Series Generator with Lip Sync Video. They generate 30-second episodes, then replace the audio track while the visual identity stays locked.

Model comparison table

Where to start

Open the Reference to Video page, upload one reference image, and run a four-second test clip. That single experiment shows how the embedding travels across models.

FAQ

What file formats work best for the first reference image? PNG or high-quality JPG at exactly 1024 by 1024 pixels gives the cleanest embedding. Lower resolution files lose fine detail in eyes and hair.

Can I change clothing without breaking the face? Yes. Add a new clothing description in the prompt while keeping the same reference embedding. The model updates only the requested layers.

How long does a character remain available? Reference embeddings stay in your account for 90 days after the last use. You can download the embedding file to keep it longer.

Does lip sync affect visual consistency? Lip Sync Video reads the same embedding file so mouth movements do not shift facial structure. The rest of the face stays fixed.

What happens if I use two different reference photos? The system merges the embeddings with a weighted average. Results vary; most users keep one primary reference for best stability.

Selecting and Preparing Reference Images

Start with a single front-facing photo taken under even lighting. Avoid hats, heavy shadows, or extreme angles because they reduce embedding accuracy when the file travels into Seedance 2.0 or Veo 3.1. Crop tightly around the head and shoulders so the 1024 by 1024 pixel canvas contains minimal background. If the source photo shows the character in motion, freeze the frame at a neutral expression before upload; this gives the embedding cleaner data for later lip-sync jobs.

Add a short text note that records exact hair length, eye color, and any permanent accessories such as glasses or earrings. The system stores these notes alongside the embedding, so future prompts only need to describe changes rather than repeat the full identity. Test the reference first on the Reference to Video page with a four-second static pose before moving to longer clips.

Managing Multiple Characters in One Project

When a scene requires two or more characters, upload each reference separately and label the embeddings clearly in your account. The platform lets you attach up to four distinct embeddings to a single job. In the prompt field, use bracketed tags such as [char1] and [char2] so the model knows which embedding controls which figure. This tagging prevents accidental face swaps during group shots.

For series work, keep a running list of every character used in an episode. Export the embedding files after each session and store them in a project folder on your computer. If you later reopen the Series Generator, you can reload the saved files instead of re-uploading photos. This practice keeps visual continuity even when team members rotate on the same show.

Troubleshooting Identity Drift

If a generated clip shows slight changes in jawline or eye spacing, reduce the motion strength slider by 20 percent on the next run. Strong motion values sometimes override the embedding in models like Kling 3.0. Another fix is to add the phrase “exact same face as reference” at the end of the prompt; the extra tokens reinforce the stored data without changing the scene description.

When drift appears across an entire sequence, regenerate the first shot with the original embedding and use that output as the new reference for subsequent shots. The Image to Video workflow accepts both the original photo and the freshly generated frame, giving the model two anchor points instead of one.

Exporting and Archiving Embeddings

After a character has been used in at least three generations, download the embedding file from the account dashboard. The file carries a .flx extension and contains the visual data plus the short text notes you entered. Store these files with version numbers so you can revert if a later update changes model behavior.

Embeddings expire from active storage after 90 days of inactivity, but the downloaded .flx file remains usable indefinitely when re-uploaded. For long-running series, schedule a monthly export of every active character so the library survives platform changes or account resets. Re-importing an old embedding into Wan 2.7 or Veo 3.1 restores the original face without needing a new reference photo.

Reference Image Quality Checklist

Use this structured list before uploading any new reference. Verify resolution first—exactly 1024 by 1024 pixels avoids downsampling artifacts inside Seedance 2.0. Check lighting next: even, soft sources from two angles produce embeddings that travel cleanly into Veo 3.1 without added shadows on cheekbones. Crop so the head occupies 70 percent of the frame and shoulders remain visible; this ratio supplies the model with enough body proportion data for later Image to Video jobs.

Inspect for accessories that must stay permanent. Note glasses frames, earrings, or scars in the accompanying text field so the embedding records them as fixed layers. Remove transient items such as scarves or hats unless they define the character. Test the prepared file with a four-second neutral-pose clip on the Reference to Video page; if eye spacing drifts more than two pixels, recapture the source photo under different light.

Store the approved reference in a dedicated folder labeled by character name and date. Keep the original high-resolution capture alongside the resized version so you can regenerate the embedding if platform defaults change.

Multi-Episode Narrative Planning

Plan an entire season arc before generating the first clip. List every scene that features each character and note required clothing states or age progression markers. Export the base embedding after the pilot episode, then create numbered variants for each new outfit by uploading the base file plus a short clothing description only.

Chain outputs through the Series Generator by loading the same tagged embeddings across episodes. When switching models mid-season, carry the .flx file rather than the original photo; this preserves the exact weighting that Seedance 2.0 learned during the first pass. Schedule a mid-season export of all active embeddings so any team member can reload the identical visual identity without re-photographing actors.

Break longer scripts into eight-second segments that end on a held pose. The final frame of one segment becomes the starting reference for the next, reducing cumulative drift across 30-minute runtimes.

Cross-Model Prompt Templates

Keep a small library of reusable prompt suffixes that reinforce identity when models change. Append “maintain exact facial structure and body proportions from reference embedding” to every text field. For environment shifts, add “background changes only; character lighting matches reference” so Wan 2.7 does not reinterpret skin tones.

When moving a character from Kling 3.0 to Veo 3.1, insert the phrase “preserve temporal attention match to stored embedding” in the first prompt of the new session. This single line restores consistency scores to within three percentage points of the original model.

Document each template version with the date and model used. Revisit the notes after any platform update so the suffix list stays current.

Team Handoff Procedures

Create a shared project folder that contains the current .flx files, prompt templates, and a one-page scene list. Label each embedding with the last episode number and clothing state. New team members import the files directly into their account dashboard instead of starting from reference photos.

Run a joint four-second test clip whenever a new member joins. Compare the output frame against the approved still; if jawline variance exceeds the tolerance noted in the checklist, regenerate the embedding from the original source before continuing production.

Archive completed episodes by exporting fresh embeddings at wrap. Store them with episode numbers so future seasons can reload the precise visual state without re-creating references from scratch.