Lawula Izithenjwa ze-Qwen Image 2.1 Nemishwana Yemiyalelo ku-ComfyUI

Lokho okwenziwa yi-ComfyUI-qwen_img_2_1_enhancer

ComfyUI-qwen_img_2_1_enhancer yengeza ama-node amabili okuqhuba ama-workflow endabuko e-Qwen Image 2.1 ku-ComfyUI. Lawa ma-node akuvumela ukuthi ulawule ukubaluleka kokunakwa kwezithombe eziyireferensi ngazinye futhi ugcizelele noma unciphise imishwana ekhethiwe ye-prompt.

Sebenzisa ama-node ngokuzimela noma uwahlanganise ku-workflow efanayo. Alungisa amaphuzu okunakwa ngaphakathi kwe-transformer endabuko ye-Qwen Image 2.1, kunokuba aphinde alinganise ama-text embedding, ahlanganise amaphikseli ezithombe eziyireferensi, noma aguqule izithombe zomthombo.

• Qwen Image 2.1 Reference Strength: Ishintsha ukubaluleka kokunakwa kwesithombe esikhethiwe esiyireferensi.
• Qwen Image 2.1 Edit — Phrase Weights: Yengeza i-syntax yemishwana enezisindo kuma-prompt amahle namabi, bese ikhiqiza okuphumayo kwe-conditioning ne-latent.

Izidingo Nokufakwa

Udinga i-ComfyUI enokusekelwa kwendabuko kwe-Qwen Image 2.1, kanye nemodeli ye-Qwen Image 2.1 esebenzayo, i-text encoder, ne-VAE setup. Le custom node ayidingi amaphakheji e-Python engeziwe noma ukuhlelwa kwamafayela ayisisekelo e-ComfyUI.

1. Vula i-terminal bese uya kufolda yama-custom node ye-ComfyUI:

cd ComfyUI/custom_nodes

1. Clone i-repository:

git clone https://github.com/capitan01R/ComfyUI-qwen_img_2_1_enhancer.git

Qala kabusha i-ComfyUI ngemva kokufaka noma ukubuyekeza i-node. Ngemva kokuqala kabusha, sesha imenyu yama-node ukuze uthole ama-node amabili e-Qwen Image 2.1 enhancer.

Lungisa Amandla Esithombe Esiyireferensi

I-node ye-Qwen Image 2.1 Reference Strength ilawula ukunakwa kwesithombe esisodwa esixhunyiwe esiyireferensi. Yengeze ngemva kwemodeli yakho nama-LoRA loader ongawakhetha, bese uxhuma okukhiphayo kwayo kwe-MODEL ku-sampler.

Load Diffusion Model -> LoRA loaders (optional) -> Reference Strength -> sampler

Xhuma izithombe eziyireferensi ne-VAE ku-node yendabuko ye-Text Encode Qwen Image 2.1 noma ku-encoder ye-Phrase Weights efakiwe. Ukukhethwa kwereferensi kulandela ukuhleleka kwezithombe ezixhunywe ku-encoder. Okokufaka kwezithombe okungenalutho kuyeqiwa.

Okokufaka kwe-Reference Strength

• model: Imodeli ye-Qwen Image 2.1, kuhlanganise nanoma yimaphi ama-LoRA asetshenzisiwe.
• reference_index: Ireferensi okufanele ilungiswe. I-1 ikhetha isithombe sokuqala esixhunyiwe, i-2 ikhetha esesibili, njalonjalo.
• strength: Isiphindaphindi samathuba okunakwa. Inani elizenzakalelayo ngu-1.2.

Ukuhumusha Amanani e-Strength

• 1.0 igcina isisindo semvelo sereferensi.
• Amanani angaphezu kwe-1.0 akhulisa ukubaluleka kwereferensi ekhethiwe.
• Amanani aphakathi kwe-0.0 ne-1.0 anciphisa ukubaluleka kwayo.
• 0.0 ivimba ukunakwa okuqondile kwamathokheni e-latent ereferensi ekhethiwe, nakuba ulwazi ngesithombe lungase luhlale kwenye indawo ku-conditioning.

Ukulungiswa kusebenza kuso sonke isithombe esiyireferensi. Ukuze ulawule amareferensi amaningi, xhumanisa ama-node amaningi e-Reference Strength. Uma inkomba efanayo iphindwa, isethingi yakamuva ithatha indawo ye-strength yangaphambilini. Ukusetha inkomba ku-1.0 kususa ukulungiswa kwayo.

Sebenzisa Imishwana Ye-Prompt Enesisindo

Sebenzisa i-Qwen Image 2.1 Edit — Phrase Weights esikhundleni se-node yendabuko ye-Text Encode Qwen Image 2.1. Xhuma i-Qwen Image 2.1 CLIP yendabuko, izithombe zakho eziyireferensi, ne-VAE. Bese uthumela okukhiphayo kwe-node kwe-positive, negative, ne-latent ku-sampler.

I-encoder ayidingi uxhumano lwe-MODEL. Imodeli ilandela indlela yayo evamile eya ku-sampler.

I-Syntax Yezisindo Zemishwana

Faka igama noma umshwana kubakaki, engeza ikholoni, bese ucacisa isisindo esingesona esinegethivu:

Add (warm lighting:1.3) and (soft shadows:1.1).

• 1.0 isebenzisa isisindo semvelo somshwana.
• Amanani angaphezu kwe-1.0 akhulisa ukubaluleka komshwana.
• Amanani aphakathi kwe-0.0 ne-1.0 anciphisa ukubaluleka komshwana.
• 0.0 icindezela kakhulu ukunakwa kwamathokheni omshwana okhethiwe.

Ama-prompt amahle namabi anezisindo ezizimele. Ukuvela okumakiwe kuphela okuthola ukulungiswa, ngakho sebenzisa amagama noma imishwana ephelele ezigabeni ezihlukene, ezingafakwanga ngaphakathi komunye nomunye.

Ukuxhumana Nokuphumayo kwe-Phrase Weights

• clip: I-text encoder yendabuko ye-Qwen Image 2.1 elayishwe ngohlobo lwe-CLIP qwen_image.
• prompt: I-positive prompt, kuhlanganise nemishwana enesisindo ongayikhetha.
• negative_prompt: I-negative prompt, enemishwana yayo enesisindo ongayikhetha.
• vae: I-VAE yendabuko ye-Qwen Image 2.1. Iyadingeka uma uhlanganisa le node ne-Reference Strength.
• resolution: Usayizi okuqondiswe kuwo lapho kushintshwa izithombe eziyireferensi. Okuzenzakalelayo ngu-1024; ukuyisetha ku-0 kugcina amareferensi eseduze nobukhulu bawo bokuqala, asondezwe abe yiziphindaphindo ze-32.
• images: Izithombe eziyireferensi. Amasokhethi ezithombe engeziwe avela njengoba kudingeka.
• positive kanye ne-negative: Okuphumayo kwe-conditioning kwe-sampler.
• latent: I-latent engenalutho elingana nereferensi yokuqala ngemva kokushintshwa kosayizi.
• phrase_report: Umbhalo we-JSON ongawukhetha obonisa ama-prompt ahlanzekile, izisindo, nemigqa yamathokheni ekhethiwe.

Sebenzisa Womabili Ama-Node Ndawonye

Ku-workflow elawula kokubili izithombe eziyireferensi nemishwana ye-prompt, dlulisa imodeli ku-Reference Strength bese uthumela okuphumayo kwe-conditioning ne-latent okuvela ku-Phrase Weights ku-sampler.

MODEL -> Reference Strength -------------------------> sampler MODEL

CLIP + VAE + images -> Phrase Weights
                           |------------------------> sampler positive
                           |------------------------> sampler negative
                           `------------------------> sampler latent_image

Uma uqhathanisa imiphumela, shintsha isilawuli esisodwa ngesikhathi ngenkathi ugcina ezinye izilungiselelo zingashintshiwe. Lokhu kwenza kube lula ukunquma ukuthi umehluko obonakalayo ubangelwa yini ukubaluleka kwereferensi noma ukubaluleka kwebinzana.

Amathiphu Athuthukile Nemikhawulo

Omabili ama-node asebenzisa ukulungiswa kwezibalo kokunaka. Ku-weight ephozithivu w, amaphuzu okunaka akhethiwe athola log(w) ngaphambi kwe-softmax. Lokhu kushintsha ukubaluleka kokunaka okuhlobene ngenkathi kugcinwa i-normalization.

• Amandla noma i-phrase weight engu-2.0 akuqinisekisi umphumela obonakalayo ophindwe kabili.
• Amanani amakhulu kakhulu angabusa eminye imiyalelo noma amareferensi futhi anciphise ukuguquguquka kokuhlela.
• Reference Strength ikhetha okhiye bezithombe zereferensi, kanti Phrase Weights ikhetha okhiye bombhalo abangamabinzana amakwe.
• Ukugcinwa kwesandulelo somdabu kuyalondolozwa.
• Ama-backend okunaka kumele asekele ama-additive mask.
• Ukumatanisa amabinzana kudinga izingcezu ezigcwele ze-tokenizer futhi akusekeli ama-embedding e-textual-inversion.
• Ukushintsha imigqa yamathokheni e-conditioning ngemva kokufakwa kwe-phrase encoding kungaphazamisa ukumataniswa kwebinzana.

Ukuze ulungise ngendlela ephathekayo, qala ngenani lomdabu elingu-1.0, wenze izinguquko ezincane, bese uhlola i-phrase_report ongayikhetha uma udinga ukuqinisekisa ukuthi yimaphi amathokheni e-prompt akhethiwe.

Isiphetho

ComfyUI-qwen_img_2_1_enhancer inikeza ukulawula okugxilile kokuhlelwa kwe-Qwen Image 2.1 ngaphandle kokushintsha amafayela ayisisekelo e-ComfyUI. Faka i-custom node, sebenzisa i-Reference Strength ukubeka phambili izithombe ngazinye, sebenzisa i-Phrase Weights ukugcizelela noma ukunciphisa amabinzana e-prompt, bese uhlanganisa womabili ama-node uma udinga ukulawula okuhlelekile kwamareferensi abonakalayo nemiyalelo yombhalo.

Iphrojekthi ikhishwa ngaphansi kwe-MIT License.