Kandinsky 6 and Wan 2.2: AI Influencer Videos with Voice
I generated 15 five-second AI influencer clips locally on an RTX 4090. Kandinsky 6 adds Polish speech; Wan 2.2 makes silent video.
👁 115 reads
I created five AI influencers as images (see the AI influencer images test), and now I’ve brought them to life. Each got a five-second vertical story in which she says one sentence in Polish to a phone. I generated the clips locally on an RTX 4090 with three models: Kandinsky 6.0 Lite and Pro (image and audio together) and Wan 2.2 5B (image only). The movement and faces look like real phone-shot stories. Polish speech is still a weak point: Whisper recognized an average of 65% of the words with the best model.
At a glance, October 17, 2026
- Input: selfie made with FLUX.2 Klein 4B; output: 5 seconds, vertical 9:16, 24 fps.
- Kandinsky 6.0 Lite: about 1.5-2.5 minutes per clip; Polish speech averaged 57% of words (range: 30-88%).
- Kandinsky 6.0 Pro (int8): about 6-7 minutes per clip; Polish speech averaged 65% of words.
- Wan 2.2 5B: about 1.5 minutes per clip, no audio, but the best-looking hand movement and facial expressions.
- All three models have licenses that allow commercial use (MIT and Apache 2.0).
Series: local AI on an RTX 4090, tests from October 17, 2026
- Roundup: new local AI models (October 2026)
- AI influencer images without the plastic look: five characters, five models
- Talking AI influencers: video with Polish speech (you’re reading this)
- FLUX 3 Image Klein: 7.5 GB VRAM, 6 seconds per image
- Qwen-Image 2.1 Turbo: 4 seconds per image
- Kandinsky 6.0: video with audio and Polish speech
- YuE2 and three other models: making Polish songs locally
How I made the clips
- I used a selfie of each character generated with FLUX.2 Klein 4B, which looked the most natural in the image test and has an Apache 2.0 license.
- I wrote one sentence (7-10 words) for each character, along with a short description of what she was doing and what could be heard in the background.
- Kandinsky got the image as its first frame, a scene description with the spoken line inside
<S>...<E>tags, and a separate audio description. Wan got the same description but no audio. - Resolution: 480×864 for Kandinsky and 544×960 for Wan; 121 frames, or 5 seconds.
- I ran the speech in each clip through Whisper large-v3 and counted how many words from the line could be heard.
The video prompt looked like this (example for the guide from Zakopane). The displayed prompt below is an English translation of the original Polish test material; the test itself was run in Polish, not English:
Five characters, three models, 15 clips
For each character, the three clips appear in this order: Kandinsky Pro, Kandinsky Lite, Wan 2.2. Kandinsky clips have model-generated audio; Wan clips are silent. The lines and Whisper transcripts below are translations of the original Polish test material. The speech was tested in Polish, not English.
Kasia, a runner from Gdańsk: “Ten kilometers before breakfast. Who’s running with me tomorrow?”
Amara, a food blogger from Wrocław: “Pierogi with jollof rice today. Sounds strange, but tastes amazing!”
Halina, a gardener from Łódź: “The carrots have done well this year. Come visit my garden!”
Mei, a tech reviewer from Kraków: “I’m testing new headphones. Review on the channel tomorrow.”
Zofia, a guide from Zakopane: “The wind is crazy, but Morskie Oko is waiting. Let’s go!”
Polish speech: what you can hear
The lines in this table are English translations of the Polish lines used in the test. The models were tested in Polish; these translated examples were not retested in English.
| Test line (English translation) | Kandinsky Pro | Kandinsky Lite |
|---|---|---|
| “Ten kilometers before breakfast. Who’s running with me tomorrow?” | 67% | 56% |
| “Pierogi with jollof rice today. Sounds strange, but tastes amazing!” | 70% | 30% |
| “The carrots have done well this year. Come visit my garden!” | 62% | 38% |
| “I’m testing new headphones. Review on the channel tomorrow.” | 86% | 71% |
| “The wind is crazy, but Morskie Oko is waiting. Let’s go!” | 38% | 88% |
Pro handled the Polish sentences better on average than Lite, but not in every clip: for the guide from Zakopane, Lite scored 88% and Pro 38%. Neither model is error-free. Words with tricky Polish spellings and unusual names such as “jollof rice” are often garbled. Short, simple sentences work best: Lite delivered the line about the wind and Morskie Oko almost without errors. For Instagram stories, where you’ll likely add captions anyway, Pro is good enough. For longer speech, it’s better to record an AI voice-over separately and add it to the video.
Kandinsky 6 vs. Wan 2.2
| Model | Time per clip | Strengths | Weaknesses |
|---|---|---|---|
| Kandinsky 6.0 Pro (int8) | 6:06-7:02 min* | Best Polish speech; background audio suits the setting | Slowest; needs 64 GB RAM |
| Kandinsky 6.0 Lite | 1:24-2:32 min | Fast, natural phone-camera movement, audio | Mispronounces Polish words (57% average) |
| Wan 2.2 5B | 1:23-2:03 min | Smooth hand movement and expressions; faithful to the first frame | No audio; the face changed slightly in one clip |
* One Pro clip took 40 minutes because another program was using the GPU at the time. I measured all times while the GPU was partly occupied. With the GPU idle, earlier Kandinsky Lite clips took about 2 minutes and Pro about 6.5 minutes (see the Kandinsky 6.0 local test).
What worked and what didn’t
- The face stays close to the image. All three models kept the freckles, glasses, hairstyles and clothes from the first frame. Only the runner in Wan’s clip looks a little younger near the end.
- Phone-like movement. Slight camera shake, a hand fixing the hair, squinting in the sun. Nobody “poses” like they’re in an ad.
- Odd gestures. In Wan’s clip, the food blogger lifts a spoonful of pierogi almost to her nose. The tech reviewer barely moves in the mirror.
- Polish words. In the Lite clip, “pierogi” turned into a word I won’t quote. Listen to the whole clip before posting it.
How to make your own clips
- Make a character image using the guidance in the AI influencer images test.
- Install Kandinsky 6 in ComfyUI (instructions in the Kandinsky 6.0 local test) and open the “Image to Video+Audio” template.
- Load the image in “Load Image,” set the width to 480 and height to 864, set
enabledto false in “Beautify Prompt,” and enter a description with the spoken line inside<S>...<E>. The prompt wording should be in the language you plan to test; the Polish test in this article was not an English-language test. - Keep sentences short, 6-10 words, and avoid foreign names. Make 2-3 versions and choose one where every word is audible.
- Use Wan 2.2 5B (the “Wan 2.2 5B ti2v” template) when you want a silent clip, for example to add music.
Remember to label the content as AI-generated. I covered the rules in the labeling section of the image test.
Frequently asked questions
Can you make a talking AI influencer for free?
Yes, locally. Kandinsky 6.0 (MIT) and Wan 2.2 (Apache 2.0) are free and allow commercial use. You need an NVIDIA GPU with 24 GB VRAM. Kandinsky Pro also requires 64 GB RAM.
Which model speaks Polish best?
Kandinsky 6.0 Pro: an average of 65% of words across my five clips. Lite scored 57%. Wan 2.2 does not generate audio.
How long can the clips be?
The models are trained on five-second clips. To make a longer story, stitch together several clips, using the last frame of one as the starting frame for the next.
How long does generation take?
On an RTX 4090, it takes about 1.5 minutes (Lite or Wan) to about 7 minutes (Pro) per five-second clip.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- Kandinsky 6.0 on Hugging Face (MIT license)
- Wan 2.2 TI2V 5B on Hugging Face (Apache 2.0 license)
- My own test on an RTX 4090, October 17, 2026: 15 clips, Whisper large-v3.
