Isifundo se-Kandinsky 6.0: Khiqiza Ividiyo Nomsindo Ovumelanisiwe Nge-AI

I-Kandinsky 6.0 ingumndeni wamamodeli e-diffusion okwenza ukukhiqizwa okuvumelanisiwe kombhalo kuya kumsindo nevidiyo. I-Kandinsky 6.0 Video Lite inamapharamitha ayizigidigidi ezi-3, kanti i-Kandinsky 6.0 Video Pro inamapharamitha ayizigidigidi ezingama-29. Womabili amamodeli angakha iziqeshana zemizuzwana emihlanu ezinomsindo we-44 kHz ovumelanisiwe, okuhlanganisa ukuvumelanisa ukunyakaza kwezindebe, kumamodi ombhalo kuya kumsindo nevidiyo kanye nesithombe kuya kumsindo nevidiyo.

Imodeli ye-plugin yokwandisa ukucaca ingakhuphula ukulungiswa kokuphumayo kube yi-Full HD, noma 1920×1080. Kulesi sifundo, uzofaka iphrojekthi, ulande i-checkpoint, ukhiqize isiqeshana sakho sokuqala, ukhethe i-device preset, futhi ubuyekeze ezinye izindlela zokusebenzisa i-Kandinsky 6.0.

Lokho okunikezwa yi-Kandinsky 6.0

I-Kandinsky 6.0 yakhelwe ukukhiqiza okuvumelanisiwe kwezinhlobo eziningi kunokukhiqiza ividiyo ethule kuphela. Amamodeli ayo akhiqiza ividiyo nomsindo ndawonye, esekela izinqubo ezisekelwe embhalweni nasezithombeni.

• Ukukhiqizwa kwevidiyo nomsindo: Dala iziqeshana zemizuzwana emihlanu ezinomsindo we-44 kHz ovumelanisiwe.
• Ukuvumelanisa izindebe: Okuqukethwe okukhiqiziwe kungafaka ukunyakaza kwezindebe okuvumelanisiwe.
• Osayizi bamamodeli ababili: Khetha phakathi komndeni we-Video Lite onamapharamitha angu-3B nomndeni we-Video Pro onamapharamitha angu-29B.
• Amamodi okufaka amabili: Sebenzisa ukukhiqizwa kombhalo kuya kumsindo nevidiyo noma kwesithombe kuya kumsindo nevidiyo.
• Ukwandisa ukucaca: Sebenzisa imodeli ye-plugin yokwandisa ukucaca ukuze ukhuphule ukulungiswa kokuphumayo kube yi-Full HD.

Izidingo

Inqubo yokuqala esheshayo yendawo idinga i-NVIDIA GPU kanye ne-Python 3.13 noma 3.14. Uzodinga futhi i-uv ne-just.

Umyalo wokusetha uthola i-GPU bese ufaka i-PyTorch build efanele. Ama-GPU e-Hopper, okuhlanganisa amadivayisi e-H100 ne-H200, asebenzisa i-PyTorch ye-CUDA 13.0 ene-FlashAttention 3. Ama-GPU e-Ampere, e-Ada, kanye ne-Blackwell asetshenziswa ngabathengi asebenzisa i-PyTorch ye-CUDA 13.0 ne-SageAttention 2.2.0.

Uma usebenzisa i-GPU ye-Ampere, i-Ada, noma i-Blackwell esetshenziswa ngabathengi, i-nvcc kumele itholakale ku-PATH yakho. Ukusetha kulindele ukuthi i-CUDA_HOME ibe yi-/usr/local/cuda ngokuzenzakalelayo.

Faka i-Kandinsky 6.0

1. Clone i-repository

Clone i-repository bese uya kumkhombandlela wephrojekthi:

git clone https://github.com/kandinskylab/kandinsky-6.git
cd k6_video
just setup

Umyalo we-just setup ufunda i-GPU yakho bese ufaka okukodwa kokucushwa okusekelwayo kwe-PyTorch ne-attention.

2. Landa imodeli

Landa i-checkpoint ye-Pro distilled ngalo myalo:

just download pro-distill

Ekusebenzeni kokuqala, iphrojekthi ilanda i-Kandinsky-6.0-Pro-distill-5s ku-Hugging Face. Ngokuzenzakalelayo, izisindo zemodeli zigcinwa ngaphansi kwe-~/.cache/kandinsky/weights.

Ungashintsha le ndawo eyisisekelo usebenzisa i-environment variable ethi KANDINSKY_HOME. Uma isethiwe, izisindo zigcinwa ku-$KANDINSKY_HOME/weights.

Khiqiza ividiyo yakho yokuqala

Ngemva kokusetha nokulanda imodeli, khiqiza isiqeshana usebenzisa i-text prompt:

just generate "a cat on a mat"

I-MP4 ephumayo ibhalwa kumkhombandlela wokuphumayo one-timestamp:

$KANDINSKY_HOME/outputs/generate_<YYYY-MM-DDTHH-MM-SS>/generations/output.mp4

Uma i-KANDINSKY_HOME ingasethiwe, iphrojekthi isebenzisa i-~/.cache/kandinsky njengomkhombandlela wayo wasekhaya ozenzakalelayo. Umkhombandlela ofanayo wokukhiqiza uqukethe ne-prompt enwetshiwe ku-expanded_prompt.txt, ukucushwa kokuqalisa ku-launch.json, kanye nemikhombandlela ye-logs/ ne-profiles/.

Umyalo wokukhiqiza ulanda i-checkpoint eshiwo ekucushweni uma ingakatholakali endaweni.

Khetha i-GPU preset

Uma i-GPU yakho idinga ukucushwa okuhlukile, dlulisa i-device preset usebenzisa i---config. Isibonelo:

just generate "a cat on a mat" --config kandinsky/configs/devices/rtx-5090.yaml --out clip.mp4

Ama-device preset agcinwe ku-kandinsky/configs/devices/. Khetha i-preset ehambisana ne-hardware nomsebenzi wakho.

Amagama atholakalayo ekhathalogi yamamodeli

Imibhalo yokuqala esheshayo ibala lawa magama ekhathalogi yamamodeli:

• pro
• pro-pretrain
• lite
• lite-distill
• lite-pretrain

Isibonelo sokuqala esizenzakalelayo sisebenzisa i-pro-distill, elanda i-checkpoint ye-Pro distilled yemizuzwana emihlanu.

Sebenzisa i-Kandinsky 6.0 ne-ComfyUI

I-Kandinsky 6.0 iphinde isekele i-ComfyUI. Faka i-kandinsky6 kanye ne-kandinsky6-sr nge-ComfyUI Manager, bese uqala kabusha i-ComfyUI.

Umkhombandlela we-comfyui/ we-repository uqukethe ikhodi yomthombo yesandiso. Ukukopisha ngesandla akudingeki. Ukuze uthole imininingwane yokufaka eqondene nephrojekthi, bheka umhlahlandlela wokusetha we-ComfyUI.

Amasu okusebenza nokwenza kahle

Isikhathi sokukhiqiza sincike kakhulu emndenini wemodeli, ekulungisweni, ku-GPU, naseqhinga lokudlulisela imithwalo. Ama-benchmark wephrojekthi alinganisa isikhathi sokusebenza sesiqeshana semizuzwana emihlanu ngemva kwe-warmup; ukulayishwa kwezisindo nokufakwa kwekhodi ye-MP4 akufakiwe.

Ukucushwa kwe-hardware okubhaliwe kusebenzisa i-module offload kumasistimu e-RTX PRO 6000, A100 80 GB, kanye ne-H100 80 GB. Amakhadi asetshenziswa ngabathengi asebenzisa i-block offload.

Qaphela ukuthi ukuqaliswa nakho kuyehluka kuye nge-GPU. Amasistimu e-RTX 4090, RTX 5060 Ti, RTX 5080, RTX 5090, RTX PRO 6000, kanye ne-A100 80 GB asebenzisa i-SageAttention2++ ekuzinakekeleni kokubona. Amasistimu e-H100 asebenzisa i-FlashAttention 3.

Humusha imiphumela ye-benchmark

Izikhathi ezibikiwe zisezuzwaneni zevidiyo yemizuzwana emihlanu esebenzisa imodeli engakancishiswa. Ukucushwa okubaliwe kufaka i-Lite SD, i-Lite HD, i-Lite Full HD, i-Pro SD, i-Pro HD, kanye ne-Pro Full HD. Ngokuvamile, ukucushwa kwe-Lite kudinga isikhathi esincane kunokwe-Pro, kuyilapho ukwenziwa kwe-Full HD kudinga isikhathi esiningi kunokwe-SD noma HD ku-hardware efanayo.

Ukuze wenze izivivinyo zendawo ezisheshayo, qala nge-checkpoint ye-Lite noma encishisiwe kanye ne-device preset efanele. Sebenzisa ukucushwa kwe-Pro noma kwe-Full HD lapho umthamo owengeziwe wemodeli noma ukulungiswa kubalulekile kumphumela wakho.

Indlela yokusebenza esebenzayo

1. Qinisekisa ukuthi isistimu yakho ine-NVIDIA GPU kanye ne-Python 3.13 noma 3.14.
2. Faka i-uv kanye ne-just.
3. Clone i-repository bese usebenzisa i-just setup.
4. Landa i-checkpoint nge-just download pro-distill noma elinye igama lekhathalogi elisekelwayo.
5. Qalisa i-just generate nge-prompt echazayo.
6. Thola i-MP4 ekhiqiziwe kufolda yomphumela enesitembu sesikhathi.
7. Buyekeza i-expanded_prompt.txt, i-launch.json, ama-log, kanye nama-profile lapho udinga ukuhlola khona i-run.
8. Shintshela ku-device preset noma ku-ComfyUI lapho indlela yakho yokusebenza idinga izilungiselelo ezihlukile ze-hardware noma ukulawula okusekelwe kuma-node.

Isiphetho

I-Kandinsky 6.0 inikeza ukwenziwa kwendawo kombhalo kuya kumsindo-vidiyo kanye nesithombe kuya kumsindo-vidiyo, okunomsindo ovumelanisiwe, ukusekelwa kwe-lip-sync, kanye ne-super-resolution ongayikhetha. Indlela elula ukufaka okuncikile nge-just setup, ulande i-checkpoint, bese usebenzisa i-prompt nge-just generate. Ukuze uthole ukulawula okwengeziwe, sebenzisa ama-device preset, hlola amafayela e-run akhiqiziwe, noma faka izandiso ze-Kandinsky 6.0 nge-ComfyUI Manager.

Ukuze uthole olunye ulwazi, vakashela iwebhusayithi ye-KandinskyLab, imibhalo ye-Diffusers, noma idemo ye-Hugging Face.