MOSS-Image Uyasebenza

I-MOSS-Image iyimodeli yokukhiqiza izithombe ngombhalo ekhishwe ngomthombo ovulekile ngokubambisana yi-Mosi Intelligence ne-OpenMOSS. Inomgogodla we-DiT onamapharamitha angu-5B, ihlanganiswe ne-encoder yombhalo ye-Qwen3.5-9B efriziwe kanye ne-FLUX.2 VAE, kanti izisindo zayo zisebenzisa i-Apache-2.0. I-MOSS-Image igcizelela ukususa izithombe ezikhiqizwe yi-AI kudatha yokuqeqesha ukuze kuncishiswe ukubukeka kwe-AI, isekele ukunikezwa kombhalo omude ngesiShayina nangesiNgisi, ukulawulwa kwesakhiwo nge-JSON Layout/Bounding Box, kanye nama-prompt afinyelela kuma-tokens angu-4096. I-MOSS-Image ikhuthaza i-ecosystem yomphakathi ehlanganisa ukusheshiswa kwezinyathelo ezingu-8/4, i-quantization ye-INT8/GGUF kanye ne-LoRA yezitayela eziningi.

Ukukhiqiza izithombe ngolimi lwemvelo: Incazelo oyifakayo ingakhiqiza izithombe ezinezitayela ezifana nokubukeka kwangempela kwezithombe, i-anime nemifanekiso.
Ukunikezwa kombhalo wesiShayina nesiNgisi: Kwenziwe kahle ekukhiqizeni imibhalo eminyene nemide, kugcinwe ngangokunokwenzeka ukwakheka okufana nokweqa kwemigqa.
Ukulawulwa kwe-Layout komdabu: Sebenzisa i-JSON ukucacisa i-summary, items, text ne-bbox, kulawulwe izindawo zezihloko, izinto nemibhalo.
Ukusekela ama-Prompt amade: Kusekela kufika kuma-tokens angu-4096, kulungele izidingo eziyinkimbinkimbi ezihlanganisa izinto eziningi, ubudlelwano bendawo, ikhamera nesitayela.
I-ecosystem nokusetshenziswa: Ihlinzeka ngezindlela ze-Diffusers, Gradio, ComfyUI, SGLang ne-vLLM-Omni, futhi iveza ukusheshiswa, i-quantization ne-LoRA.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Ukwakheka kwemodeli: Isebenzisa i-Transformer yokusabalalisa enomfula owodwa, eqeqeshwe kusukela ekuqaleni; umgogodla we-DiT unamapharamitha angu-5B. Izimo zombhalo zikhishwa ku-Qwen3.5-9B efriziwe, kanti okuguquguqukayo kwesithombe okucashile kufakwa ikhodi futhi kukhishwe nge-FLUX.2 VAE. Okuguquguqukayo kombhalo nesithombe kumodelwa ngokuhlanganyela ngaphakathi kohlaka olulodwa lokusabalalisa.
Ukulawulwa kwesakhiwo: I-summary, caption, items, label ne-bbox kuhlanganiswa kube ama-prompt ahlelekile. Izixhumanisi zisebenzisa uhlelo lwe-pixel lwe-abs_centralized nefomethi ye-yxyx, okwenza izinto, imibhalo nezindawo zibe yizimo ezicacile, hhayi izinto ezisikiswayo ngolimi lwemvelo kuphela.
Isu ledatha: Ngaphambi kokuqeqesha kwenziwa ukususa okuphindaphindiwe nokuhlunga ikhwalithi, kususwe izithombe ezikhiqizwe yi-AI, kugcinwe ukuhlukahluka kwezithombe zangempela, futhi kumakwe izici ezifana nama-watermark nokufiphala. Lokhu kunciphisa “ukunambitheka kwe-AI” futhi kuthuthukise ubunjalo bemifanekiso.
Ukuthuthukisa umbhalo nama-prompt ezinde: Imisebenzi yokuklama efana namaphosta nezithombe ze-e-commerce iqiniswa ngokunikezwa kwemibhalo eminyene/mide kanye nokufakwa kwezimo ezifika kuma-tokens angu-4096, ukuze imiyalelo eyinkimbinkimbi, okuqukethwe kombhalo nobudlelwano bendawo kungene ngesikhathi esisodwa enqubweni yokukhiqiza.

Ukulungisa indawo: Sebenzisa i-Python 3.10+ ne-NVIDIA GPU, bese ufaka i-PyTorch ehambisana ne-CUDA.
Ukulanda ikhodi: git clone https://github.com/OpenMOSS/MOSS-Image.git bese ungena kuleyo folda ukuze wakhe indawo ebonakalayo.
Ukufaka okuncikile: Ngemva kokuvula indawo, sebenzisa pip install -r requirements.txt.
Ukulayisha imodeli: Sebenzisa i-Diffusers nge-custom_pipeline="pipeline_moss_image" kanye ne-trust_remote_code=True ukuze ulayishe izisindo ze-Base/SFT.
Ukulungisa ukunemba: Kunconywa ukuthi i-DiT isebenzise i-bfloat16, i-VAE isebenzise i-float32, bese usebenzisa pipe.to("cuda").
Ukukhiqiza izithombe okuvamile: Faka i-prompt yolimi lwemvelo, usethe ububanzi nobude, num_inference_steps=50, guidance_scale=3.0 neminye imingcele ukuze ukhiqize.
Ukukhiqiza ngesakhiwo: Bhala i-summary/items/bbox njenge-JSON, uyisebenzise njenge-prompt ukuze ulawule izindawo zemibhalo nezinto.
Ama-prompt ayinkimbinkimbi: Uma izidingo zinemininingwane, sebenzisa i-Prompt Enhancer ukuze wandise umbono ube yi-caption enemininingwane noma i-Layout ehlelekile.
Ukusebenzisa i-ComfyUI: Sesha bese ufaka i-plugin ye-MOSS-Image ku-ComfyUI Manager, bese ulayisha i-workflow yesibonelo ukuze uyiqalise.
Ukusebenzisa inkumbulo encane yevidiyo: Uma i-VRAM incane, khetha i-INT8/GGUF/MLX, ukusheshiswa kwezinyathelo ezingu-8/4, noma uvule amasu okuhlukanisa abe izingxenye/ukukhipha ekulayishweni.

Ukunikezwa kombhalo kwe-LLTB: Izinga lokumbozwa kombhalo le-Mean-6 lingu-96.3%, elokuqala phakathi kwamamodeli omthombo ovulekile anezisindo ezitholakalayo ahlolwayo.
Ukulawulwa kwesakhiwo kwe-7Bench: I-mIoU yokubeka izinto ingu-0.88, i-AUC yesakhiwo ingu-0.70.
Ukuhlolwa okuyimpumputhe kwe-RLA kokuklama kwangempela: Ngaphansi kwezimo zama-prompt amafushane, isilinganiso se-Elo esilinganiselwe singu-1698.70, sisendaweni yesibili ngemuva kwe-GPT Image 2 kuphela; eqenjini lama-prompt athuthukisiwe ngokufanayo, i-Elo ingu-1807.34, idlula i-ERNIE Image ne-SenseNova U1.

Imifanekiso yemvelo kakhulu: Izithombe ezikhiqizwe yi-AI ziyasuswa kudatha yokuqeqesha, kuncishiswe “ukunambitheka kwe-AI” nokupholishwa ngokweqile.
Ukunikezwa kombhalo okuqinile: Kwenziwe kahle ngemibhalo eminyene/mide yesiShayina nesiNgisi, kanti ukumbozwa kwe-LLTB Mean-6 kufinyelela ku-96.3%.
Isakhiwo siyalawuleka: Isekela i-JSON Layout ne-Bounding Box ukuze kucaciswe imibhalo, izinto nezindawo.
Ilungele ama-Prompt amade: Ifinyelela kuma-tokens angu-4096, ilungele izincazelo eziyinkimbinkimbi zezinto, ubudlelwano, ikhamera nezitayela.
Ilula futhi kulula ukuyishintsha: Umgogodla we-5B ulungele i-fine-tuning, kanti izisindo ze-Apache-2.0 zisheshisa i-LoRA yomphakathi namamodeli asuselwe kuyo.
Izindlela eziningi zokuqalisa: Isekela i-Diffusers, Gradio, ComfyUI, SGLang, vLLM-Omni kanye ne-API.
Isebenza kahle emisebenzini yokuklama: Igxile kumaphosta, izithombe ze-e-commerce nemisebenzi yokuhlela, kanti imiphumela yokuhlolwa okuyimpumputhe isondele noma idlula eminye yamamodeli ezentengiso.

Iwebhusayithi yephrojekthi: https://openmoss.ai/MOSS-Image/
Indawo ye-GitHub: https://github.com/OpenMOSS/MOSS-Image
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base
Umbiko wobuchwepheshe: https://github.com/OpenMOSS/MOSS-Image/blob/main/paper/MOSS_Image_Technical_Report.pdf

Ukuklama amaphosta: Sebenzisa i-JSON Layout ukulungisa izindawo zesihloko, isithombe esiyinhloko nezinto ezisekelayo, ukuze ukhiqize ngokushesha amaphosta okukhangisa anesakhiwo esilawulekayo.
Ukukhiqiza izithombe ze-e-commerce: Lawula izindawo zomkhiqizo, umbhalo wokukhangisa, ilebula lentengo nesizinda ukuze ukhiqize izithombe zokumaketha ezifanele amakhasi emininingwane nezithombe eziyinhloko.
Izithombe zokuqukethwe kwezokuxhumana: Guqula ama-Prompt amade abe okuqukethwe okungokoqobo noma okuyimifanekiso okuhambisana nesitayela somkhiqizo, okusetshenziselwa i-WeChat Official Accounts, i-Xiaohongshu nezinto zokukhangisa.
Ukudala imifanekiso ye-anime: Hlanganisa i-LoRA yezitayela eziningi ukuze ukhiqize abantu, izigcawu nemidwebo yomqondo, kusetshenziswe emisebenzini ye-otaku, izincwadi zezithombe nokuthuthukiswa kokubukwa.
Izinto zokuhlela zesiShayina nesiNgisi: Phatha imisebenzi yokuklama eyenziwe yaba ngeyasendaweni enemibhalo eminyene, njengamamenyu, amakhasi emicimbi, izitikha nemiyalelo yokupakisha.