Wan-Animate-2 – imodeli yesizukulwane esisha yokugqwayiza abalingiswa ekhishwe njenge-open source yiqembu le-Wanxiang Uyasebenza

I-Wan-Animate-2 iyimodeli yesizukulwane esisha yokugqwayiza abalingiswa ekhishwe njenge-open source yiqembu le-Wanxiang. Njengokuthuthukiswa okukhulu kwesakhiwo se-Wan-Animate, isusa ukuncika ohlakeni lokuma olucacile nakunethiwekhi eyisizayo yokukhipha ukuma, bese isebenzisa i-Diffusion Transformer (DiT) enamagatsha amabili esebenza ekugcineni kuya ekugcineni ukuze ithwebule ngokuqondile ulwazi lokunyakaza oluvela kuvidiyo eyisithenjwa. Imodeli isekela ukulawulwa kwe-engeli yokubuka okuqhutshwa umbhalo futhi inikeza uhlobo lokusakaza ngesikhathi sangempela olungafinyelela ku-24 fps.

Ukuguqula ividiyo eqondisayo ibe ukugqwayiza komlingiswa: Faka isithombe esiyisithenjwa reference.png kanye nevidiyo yethempulethi eqondisayo template.mp4, ukuze umlingiswa oyisithenjwa aphinde enze ukunyakaza nesigqi sekhamera esikwi-video eqondisayo.
Ukudluliselwa kokunyakaza ekugcineni kuya ekugcineni: I-DiT eklanywe kabusha isebenzisa ngokuqondile ividiyo eqondisayo, ngaphandle kwesikhiphi sokunyakaza esiphakathi, ngenhloso yokuthuthukisa kokubili ukwethembeka kokunyakaza nokuvumelana kobuwena.
Ukugcina ubuwena nokubukeka: I-reference image token kanye ne-reference video token zibamba iqhaza ekubalweni kwe-attention, zifake ukubukeka kwesithenjwa enqubweni yokukhiqiza futhi zinciphise izimo lapho “ukunyakaza kufana kodwa ubuso bonakele”.
Ukulawulwa okujwayelekile kombhalo: Ngaphambi kwe-inference, sebenzisa i-LLM ukukhiqiza i-caption enefomethi engaguquki, ebhala kuphela “incazelo yokubukeka komuntu + incazelo yangemuva”, ngaphandle kokuchaza ukunyakaza noma ukuqagela imizwa, bese uyisebenzisa njenge-prompt ye-Wan-Animate-2.
Ukulawulwa kwe-engeli yokubuka okuqhutshwa umbhalo: Umbhalo ungasetshenziswa ukuhlukanisa i-engeli yekhamera yokuphumayo kuvidiyo eqondisayo, okwenza kube lula ukukhiqiza “isethi efanayo yokunyakaza, kodwa izindawo/ama-engeli ekhamera ahlukene”.
Imodi ye-Base yekhwalithi ephezulu: Iskripthi esizenzakalelayo se-inference sisebenzisa izilungiselelo ezijwayelekile; isibonelo esisemthethweni ngasohlangothini lwe-Diffusers sisebenzisa num_inference_steps=40, sigxile kakhulu kwikhwalithi.
Imodi ye-Distillation enezinyathelo ezimbalwa: Isibonelo sezisindo ezi-distillwe sisebenzisa 10 steps, guidance_scale=1.0, kanye ne-Euler solver/no CFG, siqondiswe ekukhiqizeni ngokushesha nasekuqinisekiseni ukubambezeleka okuphansi.
Umsebenzi obheke ekusebenzeni ngesikhathi sangempela: Iphepha liphakamisa i-Wan-Animate-2-Lite, esebenzisa i-teacher forcing, i-error buffer, i-Self-Forcing distillation kanye ne-chunk-wise backpropagation ukuze icindezele ukubambezeleka kusondele emngceleni wesikhathi sangempela, iqondiswe ekugqwayizeni abalingiswa okusakazwayo.
Okuhlangenwe nakho kwasendaweni nokuku-inthanethi: Kunikezwa izikripthi ze-Gradio zasendaweni; i-Base isebenzisa wan_animate_2_gradio.py, kanti i-distillation isebenzisa wan_animate_2_gradio_distillation.py. I-README iphinde ikhombe ku-demo eku-inthanethi ye-ModelScope Studio.
Indawo yokungena ye-ecosystem yobunjiniyela: Izisindo zitholakala ku-HuggingFace/ModelScope, futhi isibonelo se-ID yemodeli ngu-Wan-AI/Wan2.2-Animate-2-14B. I-Diffusers ingafakwa kusuka kukhodi yomthombo ukuze ihlanganiswe; i-ComfyUI/DiffSynth-Studio zisaseTodo.
Ukukhiqiza okuhambisanayo okulungisekayo: Izilungiselelo ezizenzakalelayo zenzelwe ama-8×A800 ne-720P, futhi kuhlolwe nama-2×A800 ne-480P. Izingxenyekazi zekhompuyutha ezihlukile zidinga ukulungiswa kwe-YAML parallel configuration.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI open source
Lungisa i-hardware: Hlela kuqala i-8×A800 ne-720P ezenzakalelayo esemthethweni; kuphinde kwahlolwa i-2×A800 ne-480P. Ungacabangi ngokuzenzakalelayo ukuthi ikhadi elilodwa lezinga lomthengi lizokwazi ukuyisebenzisa.
Clone i-repository: git clone --recursive https://github.com/Wan-Video/Wan-Animate-2.git.
Setha imvelo: Python 3.11; faka torch 2.7.0 / torchvision 0.22.0 / torchaudio 2.7.0 kusuka ku-CUDA 12.6; bese ufaka requirements.txt ne-flash-attn, ekugcineni usebenzise pip install -e .
Landa izisindo: Ku-HuggingFace sebenzisa huggingface-cli download Wan-AI/Wan2.2-Animate-2-14B --local-dir ./ckpts/; ku-ModelScope sebenzisa modelscope download --model Wan-AI/Wan2.2-Animate-2-14B --local_dir ./ckpts/.
Qala ngokukhiqiza i-caption: Sebenzisa i-LLM efana ne-Qwen3.7-Plus, ulandele ifomethi yesiShayina esemthethweni futhi ubhale kuphela “incazelo yokubukeka komuntu + incazelo yangemuva”; ungabhali ukunyakaza, imizwa noma ukuhlola.
Qalisa i-Base: Ngena ku-infer, usebenzise wan_animate_2_demo.py, bese udlulisa --prompt, --refer-img-file, --refer-video-file, --config ./wan_animate_2.yaml.
Qalisa uhlobo lwe-distillation: Shintshela ku-wan_animate_2_distillation.yaml bese wengeza --sample_guide_scale 1.0 --step 10; ngasohlangothini lwe-Diffusers lokhu kuhambisana ne-guidance_scale=1.0, flow_solver="euler".
Lungisa ukucushwa kokuhambisana: Uma inani lama-GPU noma imemori lingafani nokusemthethweni, qala ngokulungisa ama-parallel configs ku-YAML, bese uphenya i-OOM noma ijubane elingajwayelekile.
Ukuhlola okusheshayo: Uma ungafuni ukuyifaka, qala uzame i-ModelScope Studio; endaweni yangakini sebenzisa wan_animate_2_gradio.py noma wan_animate_2_gradio_distillation.py.

Ukuthembeka kokunyakaza nokuvumelana kobuwena yizona zinto ezikhangayo kakhulu: ngokususa isikhiphi sokunyakaza esiphakathi, ngokombono kuncipha ukuqoqeka kwamaphutha kwesinye nesinye isigaba.
Isondelene kakhulu nochungechunge lokukhiqiza: I-caption ibekwe ngaphambili njengefomethi engaguquki, kudingeka kuchazwe ukubukeka komuntu kanye nengemuva, kungachazwa ukunyakaza, okwenza ukuphindaphinda okuzinzile kube lula.
Indlela yokubambezeleka okuphansi icacile: I-Distillation isivele ingenza inference ngezinyathelo eziyi-10; iphepha le-Lite liqondise ngqo kubantu bedijithali besikhathi sangempela nabasakazwayo.
Ifanele ukusetshenziswa kwezentengiso ngezinga eliphezulu: I-Apache-2.0 inobungane kakhulu ekuthuthukisweni kwesibili, kodwa ukusetshenziswa kwezentengiso kusadinga ukuphathwa kwamalungelo ezinto, amalungelo esithombe somuntu kanye nemithetho yezinkundla.

Iwebhusayithi esemthethweni yephrojekthi: https://humanaigc.github.io/wan-animate-2
I-repository ye-Github: https://github.com/Wan-Video/Wan-Animate-2
ModelScope: https://modelscope.cn/models/Wan-AI/Wan2.2-Animate-2-14B
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2608.06009
I-demo eku-inthanethi: https://www.modelscope.cn/studios/Wan-AI/Wan2.2-Animate

Abasakazi ababonakalayo / ukusakaza komuntu wedijithali: Indlela ye-Lite iqondiswe ngokusobala emngceleni wesikhathi sangempela nasekugqwayizeni abalingiswa okusakazwayo, ilungele amagumbi okusakaza abambezeleka kancane lapho “umuntu wangempela eshayela isithombe esibonakalayo”.
Abalingiswa bamavidiyo amafushane / amadrama amafushane: Sebenzisa ividiyo yokudlala komlingisi njenge-driving video, uyifake emfanekisweni womlingiswa noma esithombeni esiyisithenjwa. Lokhu kulungele abalingiswa befantasy, izithombe ze-IP kanye nezigcawu ezibiza kancane zokushintsha abalingisi; okubalulekile ukudluliselwa kokunyakaza ekugcineni kuya ekugcineni nokugcinwa kobuwena.
Umdanso, ama-MV kanye nokudalwa kabusha komdanso wezandla: Sebenzisa ithempulethi eyodwa yomdanso ukushayela izithombe zabalingiswa abahlukene, wakhe ukudansa kabusha kwabalingiswa abaningi, ukushintsha izingubo esithombeni esisodwa noma ama-MV anesitayela. Ukulawulwa kwe-engeli yokubuka kungakhiqiza nezinto zokuhlela “zokunyakaza okufanayo kodwa izindawo zekhamera ezahlukene”.
Okuqukethwe kwe-e-commerce nemikhiqizo: Vumela i-mascot yomkhiqizo noma imodeli ebonakalayo ukuthi ibonise ukunyakaza ngethempulethi efanayo, ukuze kusetshenziswe kumavidiyo okwethulwa kwemikhiqizo emisha, amakhasi emicimbi kanye nezinto zemithombo yezokuxhumana. Ngaphambi kokusetshenziswa kwezentengiso, kufanele kuphathwe amalungelo esithombe somuntu, amalungelo ezimpahla zokugqoka kanye nokuhambisana nemithetho yenkundla.
Izifundo / izinto zokufundisa zababhali bolwazi: Shayela isithombe sikathisha noma uthisha wekhathuni ngethempulethi yokunyakaza engaguquki, ukhiqize ngobuningi izigcawu zokuqala, zokushintsha kanye nezaziso zamaphuzu olwazi, wehlise izindleko zokuphinda kuthathwe izithombe.
Ama-NPC emidlalo / ukuhlela kusengaphambili kwezigcawu ezidlulayo: Abahleli bangasebenzisa kuqala ukudlala komuntu wangempela ukuze bashayele ngokushesha uhlobo lokuqala lomlingiswa, baqinisekise isigqi sokunyakaza, ikhamera nemizwa, bese benquma ukuthi bazongena yini ochungechungeni olusemthethweni lokuthwebula ukunyakaza nokunikezela.