Reference to Video Tutorial 2026
Fix character drift in generated video by anchoring a reference image. Follow the exact upload, model, and strength settings that keep landmark error under 3 percent.
TL;DR
Upload a 1024 by 1024 reference face to the reference-to-video tool, select Seedance 2.0 or Veo 3.1, set reference strength at 0.75 and motion at 0.65, then generate 8-second clips. This workflow keeps facial drift below 3 percent across frontier models.
The character drift problem hits at frame 12
You upload a clean reference face to a video generator and watch the eyes shift by frame 12. The mouth follows at frame 18. Credit spend rises while output quality drops. This exact mismatch drives most users to the reference-to-video workflow.
Why generic text-to-video pipelines fail
Text prompts alone cannot lock identity across time. Seedance 2.0 and Veo 3.1 both require an image anchor to hold facial structure. Without the anchor, even detailed prompts produce drift after eight seconds. Kling 3.0 reduces the issue but still needs the reference upload to stay under 4 percent landmark error.
How the reference pipeline works in practice
Flixly routes the reference image through a dedicated encoder before the diffusion pass. The system then blends the encoded face into every subsequent frame using the selected model. Users see the first stable output in 22 seconds on average when credits are pre-loaded.
Start at the Reference to Video page. Upload a 1024 by 1024 PNG face shot with neutral lighting. Select Seedance 2.0 from the model dropdown. Set motion strength to 0.65 and duration to 8 seconds. Add a short prompt such as "slow pan right, daylight street". Submit the job.
Model comparison for reference consistency
Different frontier models handle the same reference at different fidelity levels.
| Model | Max Duration | Landmark Drift | Credit Cost | Best Use Case |
|---|---|---|---|---|
| Seedance 2.0 | 12 s | 2.1 % | 18 | Talking head close-ups |
| Kling 3.0 | 10 s | 3.4 % | 15 | Full-body walk cycles |
| Veo 3.1 | 8 s | 1.8 % | 22 | Complex background motion |
| Wan 2.7 | 6 s | 4.9 % | 12 | Quick test renders |
The table shows Seedance 2.0 keeps drift lowest for mid-length clips while Veo 3.1 excels when background elements must move independently.
Step-by-step reference upload process
- Open the dashboard and navigate directly to the reference tool. Confirm your account has at least 20 credits remaining.
- Drag a square-cropped reference photo into the upload zone. The system auto-detects 1024-pixel width and rejects files below 512 pixels.
- Choose the target model from the selector. Seedance 2.0 appears first in the list for new accounts.
- Adjust the reference strength slider to 0.75. Values above 0.85 flatten motion too much.
- Enter the motion prompt in the text field. Keep it under 18 words for best token alignment.
- Pick output resolution and frame rate. 720p at 24 fps matches most social export needs.
- Click generate and monitor the progress bar. Download the MP4 once the job reaches 100 percent.
- Review the result in the preview player. If eye drift appears, lower motion strength by 0.15 and rerun.
Edge cases and hard limits
Reference images with heavy makeup or side lighting increase drift by 3 percent on average. Wan 2.7 cannot exceed 6 seconds without visible stretching. Voice cloning layers require a separate Lip Sync Video pass after the reference video finishes. The platform does not support 4K reference input; upscale first via AI Image Tools.
Users who need longer clips combine two 8-second reference outputs inside Video to Video. This adds 14 credits and 45 seconds of processing time.
Quick settings cheat sheet
Keep reference strength between 0.65 and 0.80 for natural head turns. Motion strength above 0.70 works only with Seedance 2.0. Always match prompt language to the model; English yields lowest error rates on Veo 3.1.
FAQ
What resolution should the reference photo be? Use exactly 1024 by 1024 pixels. Lower sizes trigger an automatic upscale that softens fine facial details.
How many seconds can I generate with one reference? Seedance 2.0 supports up to 12 seconds before drift exceeds 5 percent. Split longer scenes into multiple jobs.
Can I change clothing after the first render? Yes. Run the output through Image to Video with a new clothing reference and keep the face locked.
Does the tool support anime style references? Switch the model to Anime / Series Generator after the initial reference pass. The same 1024-pixel upload works.
Why does my output show background warping? Lower motion strength to 0.55 or switch to Veo 3.1 which handles independent background layers better.
Frequently Asked Questions
What resolution should the reference photo be?▾
Use exactly 1024 by 1024 pixels. Lower sizes trigger an automatic upscale that softens fine facial details.
How many seconds can I generate with one reference?▾
Seedance 2.0 supports up to 12 seconds before drift exceeds 5 percent. Split longer scenes into multiple jobs.
Can I change clothing after the first render?▾
Yes. Run the output through Image to Video with a new clothing reference and keep the face locked.
Does the tool support anime style references?▾
Switch the model to Anime / Series Generator after the initial reference pass. The same 1024-pixel upload works.
Why does my output show background warping?▾
Lower motion strength to 0.55 or switch to Veo 3.1 which handles independent background layers better.
