Dala Abalingiswa Abangaguquki Ngesithombe Esisodwa Sereferensi Usebenzisa i-Anima Reference Adapter

Yini i-Anima Reference Adapter?

I-Anima Reference Adapter iyisethi yama-node e-ComfyUI enziwe ngokwezifiso, esetshenziselwa ukukhiqiza isithombe se-Anima kusetshenziswa isithombe esisodwa sereferensi kanye ne-prompt echaza isigcawu kuphela. Izama ukulondoloza ukuthi ubani obonisiwe—okuhlanganisa izinwele, amehlo, izingubo, izesekeli, kanye nenye indlela yokudweba—ngenkathi ivumela i-prompt ukuthi ilawule ukuma, ingemuva, isenzo, indlela yokufreyima, kanye nokwakheka kwesithombe.

Awudingi ukufaka igama lomlingiswa, i-artist tag, noma ama-tag anemininingwane yokubukeka. Lokhu kwenza i-adapter ibe wusizo kubalingiswa bokuqala kanye nabalingiswa abaziwayo. Inguqulo yamanje ukukhishwa kokuhlola okuqeqeshwe ngabalingiswa abangu-300 ngezinyathelo ezingu-2,000.

Lesi sibonelo sichaza ukuhamba komsebenzi okuhlosiwe: ubuwena obubonakalayo buvela kureferensi, kanti i-prompt efana nokuthi full body, walking, hands in pockets, city, street, night, rain ichaza isigcawu esisha.

Izici Eziyinhloko

• Ukuqondiswa ngesithombe esisodwa: Faka isithombe esisodwa ukuze udlulisele ubuwena bomlingiswa kanye nezinye izimfanelo zesitayela.
• Ukubhala i-prompt yesigcawu kuphela: Chaza isenzo, ukuma, indawo, indlela yokufreyima, nokukhanya okufunayo esikhundleni sokuphinda uchaze ukubukeka komlingiswa.
• Ukuqondisa okuzimele: Lungisa i-text CFG kanye nokuqondiswa kwe-referensi ngokuhlukana.
• I-APG nokulungiswa kwe-rescale: Lawula ukugcwala kwemibala okweqile kumazinga aphezulu okuqondiswa kwe-referensi.
• Ukuhlanganiswa ne-ComfyUI: Iphrojekthi inikeza i-loader, i-reference-latent node, kanye ne-custom guider.
• Ukusekela abalingiswa bokuqala: Le ndlela ayincikile egameni lomlingiswa elaziwayo.

Indlela I-Adapter Esebenza Ngayo

Isithombe sereferensi silinganiswa ukuze silingane nobukhulu bokukhishwa obuceliwe, kusetshenziswa ama-white letterbox, bese sifakwa ikhodi yi-Qwen-Image VAE. Siba ifreyimu yesibili ehlanzekile ye-latent ku-timestep zero eduze kwe-target latent enomsindo.

Izixhumanisi ze-RoPE zefreyimu yereferensi zisuselwa ngaphandle kwegridi yesithombe esiqondiwe. Lokhu kunciphisa ukukopishwa okuqondaniswe namaphikseli, kuvumele ukuma nokwakheka kwesithombe okuceliwe kuhluke kureferensi. I-LoRA esezingeni le-rank-32 equkethe amapharamitha ayizigidi ezingu-45.9 ifundisa i-Anima ukuthi ifunde le frayimu eyengeziwe. I-adapter LoRA isebenza kuphela lapho kukhona ireferensi.

Ukuqondiswa kombhalo kanye nereferensi okuhlukene kukuvumela ukuba ulinganise ukulandela i-prompt ngokumelene nokudluliswa kobuwena nesitayela. I-adaptive projected guidance, noma i-APG, iyatholakala ukusiza ukugcina imibala ilawulwa lapho ukuqondiswa kwe-referensi kuphezulu.

Izidingo

Ngaphambi kokufaka ama-custom node, lungisa ukufakwa kwe-ComfyUI okusebenzayo kanye nezingxenye zemodeli ezidingwa yi-Anima:

• Anima base v1.0
• Qwen3-0.6B text encoder
• Qwen-Image VAE
• Anima Reference Adapter pilot weights

Izingxenye ze-Anima ziyatholakala kurepository ye-CircleStone Labs Anima. Ama-adapter weights ayatholakala kurepository ye-Anima Reference Adapter ku-Hugging Face.

Ukufaka

1. Faka ama-custom node

Vula i-terminal, ungene ku-directory ye-custom_nodes ye-ComfyUI, bese u-clone i-repository:

cd ComfyUI/custom_nodes
git clone https://github.com/ElectronGirl/ComfyUI-Anima-Ref-Adapter.git

2. Dawuniloda ama-adapter weights

Dawuniloda i-anima_ref_adapter_pilot_r32.safetensors bese uyibeka kule directory elandelayo:

ComfyUI/models/loras/anima_ref_adapter_pilot_r32.safetensors

3. Faka noma weqe i-Patch Sage Attention

Zombili izibonelo zokuhamba komsebenzi zisebenzisa i-Patch Sage Attention node evela ku-ComfyUI-KJNodes. Faka i-KJNodes uma usebenzisa i-SageAttention. Kungenjalo, khetha leyo node ekuhambeni komsebenzi bese uyakweqa ngo-Ctrl+B.

4. Qalisa kabusha i-ComfyUI

Qalisa kabusha iseva ye-ComfyUI ngemva kokufaka ama-node nama-weights. Layisha kabusha isikhombimsebenzisi sesiphequluli ngo-Ctrl+F5 ukuze ama-node amasha abhaliswe.

Khiqiza Isithombe Sakho Sokuqala

1. Vula ukuhamba komsebenzi okuyisisekelo

Layisha i-workflows/anima_ref_adapter.json evela ku-repository oyiclone ku-ComfyUI. Okunye ukuhamba komsebenzi, i-workflows/anima_ref_adapter_pag.json, kufaka i-Anima Safe PAG futhi kuyasiza ngemva kokuba ukusetha okuyisisekelo sekusebenza.

2. Layisha isithombe sereferensi

Khetha isithombe esicacile somlingiswa ofuna ukudlulisa ubuwena nesitayela sakhe. I-node ye-Anima Ref Latent ilinganisa isithombe ukuze singene kubukhulu bokukhishwa obukhethiwe, ingeze ama-white letterbox lapho kudingeka, bese isifaka ikhodi nge-VAE. Iphinde idale i-target latent engenalutho kanye ne-reference encode enobukhulu obuyingxenye, ukuze kusetshenziswe ekuhambeni komsebenzi okuhambisana ne-multi-resolution.

3. Bhala i-prompt yesigcawu kuphela

Sebenzisa amagama ekhwalithi, ubulili bomlingiswa, solo, kanye nama-tag esigcawu. Ungafaki igama lomlingiswa, igama lomculi, noma incazelo yokubukeka. Isakhiwo esijwayelekile se-prompt yilesi:

masterpiece, best quality, score_7, newest, highres, absurdres, safe, <gender>, solo, <scene>

Isibonelo, ukuze ubeke umlingiswa wereferensi ogwini ngesikhathi sokushona kwelanga, sebenzisa:

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, cowboy shot, looking back, smile, beach, ocean, sunset

Eminye imibono yama-prompt egxile esigcawini kuphela ihlanganisa:

• upper body, sitting, holding cup, drinking, cafe, indoors, window, sunlight
• portrait, from side, looking afar, cherry blossoms, petals
• full body, standing, snow, forest, winter
• upper body, reading, holding book, library, bookshelf

4. Sebenzisa izilungiselelo ezinconyiwe

Kwimodeli eyisisekelo, qala ngalezi zilinganiso:

• Compute dtype: bf16 nge-ModelComputeDtype
• Sampler: Euler
• Steps: 28 kuya ku-30
• Scheduler: simple
• Text CFG: 4.5
• Reference guidance: 4.5 kuya ku-6
• Correction: apg, noma rescale nge-phi 0.5
• Reference offset: Gcina i-ref_offset_h enanini layo lokuqeqeshwa elingu-1.0

Sebenzisa i-bf16 esikhundleni se-fp16. Ngokwemibhalo yephrojekthi, i-Anima iba ne-overflow ku-fp16.

5. Faka i-workflow emgqeni

Qinisekisa ukuthi imodeli, i-VAE, i-text encoder, isithombe sereferensi, izisindo ze-adapter, kanye ne-prompt kuxhunywe kahle. Faka i-prompt emgqeni ukuze ukhiqize isithombe okuqondwe kuso.

Ukuqonda ama-Custom Node

Anima Ref Adapter Loader

Le node ixhuma i-adapter LoRA ku-MODEL ye-Anima. Gcina i-ref_offset_h iku-1.0, okuyinani elisetshenziswe ngesikhathi sokuqeqeshwa.

Anima Ref Latent

Le node ilungiselela ama-latent ereferensi nawethagethi. Ilinganisa futhi ibeke i-reference ngendlela ye-letterbox, iyifake nge-VAE, idale i-half-size encode, bese ibuyisa i-latent engenalutho ngosayizi wokuphumayo okhethiwe.

Anima Ref Guider

Le node ikhiqiza i-GUIDER esetshenziswa yi-SamplerCustomAdvanced. Inikeza izilawuli ezihlukene ze-text cfg ne-ref_guidance, kanye nokulungiswa kwe-APG noma kwe-rescale. Iphinde isebenzise ama-patch e-post-CFG afana ne-PAG.

Amathiphu Okulungisa Athuthukile

Linganisela ubunikazi nokulandela i-prompt

I-reference guidance ilawula ukuthi umphumela ulandela kangakanani i-reference. Inani elingu-1 linikeza i-conditioning ebuthaka nengabizi kakhulu. Amanani asuka ku-4.5 kuya ku-6 ngokuvamile agcina imininingwane eminingi yomlingiswa nesitayela, kodwa ukwandisa inani kungadala nokugcwala kombala okunamandla. I-APG iwukulungiswa okunconyiwe uma kusetshenziswa i-guidance ephezulu.

Sebenzisa ama-LoRA ozikhethela wona ngokulandelana okufanele

Ama-LoRA ajwayelekile kufanele axhunywe ngaphambi kwe-Anima Ref Adapter Loader. Izibonelo zegalari zisebenzise i-Anima context-detailer LoRA cishe ku-0.3 kanye ne-base-slider LoRA cishe ku-0.6. Lezi zengezo ziyizinketho, akuzona izimfuneko zokufakwa.

Engeza i-Anima Safe PAG

Ukuze wakhe igraph ethuthuke kakhulu, vula workflows/anima_ref_adapter_pag.json. Ukucushwa kwegalari kusebenzise i-Anima Safe PAG ku-block 18, nge-scale engu-1.3 esikhathini esisuka ku-0.2 kuya ku-0.9.

Sebenzisa i-Anima turbo LoRA ngokucophelela

I-Anima turbo LoRA nayo isebenza ne-adapter. Yilayishe ngaphambi kwe-adapter loader, bese usebenzisa izilungiselelo ezijwayelekile ze-step ne-CFG ze-turbo LoRA esikhundleni sezincomo zemodeli eyisisekelo. I-reference adapter uqobo yaqeqeshwa ku-Anima base v1.0.

Khetha i-sampling enokulungiswa okukodwa

I-Euler enama-step angu-28 kuya ku-30 kanye ne-simple scheduler iyindawo yokuqala enconyiwe. I-README ibika ukuthi isigaba se-half-resolution se-SamplerSPEED asithembekile ngoba i-pilot adapter ayiqeqeshelwanga lolo hlelo lwe-progressive sampling.

Thuthukisa i-reference input

Sebenzisa i-reference lapho izici ezibalulekile zobunikazi zibonakala khona, kodwa lindela ukuthi i-pose nokwakheka okukhiqiziwe kulandele i-prompt kunokuba kulandele i-reference. Uma i-accessory noma iphethini elincane linyamalala, ukwandisa i-reference guidance kungasiza, nakuba imodeli ye-pilot inemikhawulo eyaziwayo emininingwaneni emincane.

Imikhawulo Yamanje

• Ubunikazi bomlingiswa budluliselwa ngokuthembeke kakhulu kunesitayela sokudweba.
• Izitayela ezipendwe kakhulu zingadluliselwa kabi.
• Amaphethini amancane, izibazi, amacici, namathoni esikhumba angajwayelekile kuvame ukulahleka.
• Izingubo ezinkulu zingashintsha umbala.
• Umlingiswa osetshenziswe njenge-reference kwesinye isikhathi angaphinde avele njengento ngaphakathi kwesigcawu.
• I-adapter yaqeqeshwa kuphela ezithombeni ezikhiqizwe yi-Anima futhi izuza ukuchema kwemodeli eyisisekelo.
• Ama-reference ahlukene omlingiswa nesitayela awatholakali kulokhu kukhishwa kwe-pilot.

I-roadmap ihlanganisa izithombe ezihlukene zomlingiswa nesitayela, idathasethi enkulu enabalingiswa abasha abangu-2,000 ngezitayela zabaculi ezintathu, ukudluliswa okungcono kwesitayela esipendwe kakhulu nemininingwane emincane, kanye nama-pair okuqeqeshwa e-half-resolution ama-progressive sampler.

Ilayisensi

Ikhodi ye-ComfyUI node isatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0. Izisindo ze-adapter ziyinzalo ye-CircleStone Model futhi zisebenzisa i-CircleStone Non-Commercial License. Buyekeza ilayisensi yezisindo ngaphambi kokusebenzisa umsebenzi okhiqiziwe noma uhlanganise i-adapter kuphrojekthi. I-repository ithi le phrojekthi ayihlangene ne-CircleStone Labs futhi ayigunyazwanga yiyo.

Isiphetho

I-Anima Reference Adapter inikeza indlela esebenzayo yokuyisa umlingiswa ezigcawini ezintsha ngaphandle kokubhala omaka bobuwena, bomculi, noma bokubukeka. Qala ngohlelo oluyisisekelo, sebenzisa i-bf16, isampula ye-Euler, i-CFG engu-4.5, kanye nokuqondisa kwereferensi okuphakathi kuka-4.5 no-6. Uma uhlelo oluyinhloko selusebenza, engeza i-APG, i-Anima Safe PAG, noma ama-LoRA ozikhethela wona, ugcine ama-LoRA ajwayelekile engaphambi kwesilayishi se-adapter.