Stable Signer:
Hierarchical Sign Language Generative Model

Sen Fang1, Yalin Feng2, *, Hongbin Zhong3, Yanxin Zhang4, Dimitris N. Metaxas1
1Rutgers University, 2Nanyang Technological University, 3Georgia Institute of Technology, 4University of Wisconsin-Madison
* Equal Contribution
ICCV 2025 Demo Track †
Paper arXiv Data

Abstract

Sign Language Production (SLP) is the process of converting the complex input text into a real video. Most previous works focused on the Text2Gloss, Gloss2Pose, Pose2Vid stages, and some concentrated on Prompt2Gloss and Text2Avatar stages. However, this field has made slow progress due to the inaccuracy of text conversion, pose generation, and the rendering of poses into real human videos in these stages, resulting in gradually accumulating errors. Therefore, in this paper, we streamline the traditional redundant structure, simplify and optimize the task objective, and design a new sign language generative model called Stable Signer. It redefines the SLP task as a hierarchical generation end-to-end task that only includes text understanding (Prompt2Gloss, Text2Gloss) and Pose2Vid, and executes text understanding through our proposed new Sign Language Understanding Linker called SLUL, and generates hand gestures through the named SLP-MoE hand gesture rendering expert block to end-to-end generate high-quality and multi-style sign language videos. SLUL is trained using the newly developed Semantic-Aware Gloss Masking Loss (SAGM Loss). Its performance has improved by 48.6% compared to the current SOTA generation methods, which is a significant increase in the SLP field.

More Information

We present some latest results of our newly developed SLP model. Our new work aims to combine the most stable and fastest Text2Pose sign language model with state-of-the-art Pose2Video model to help more disadvantaged groups.

We are the team with the fastest progress and best results in the field of multilingual generation. While the United Nations recognizes sign language as one of its official languages, the deaf community still faces significant communication challenges. Currently, mainstream academia can only generate German Sign Language and American Sign Language, overlooking a broader population worldwide. We plan to target more than eight sign languages for our generation system, which would cover over fifty countries.

If you have any regular or academic cooperation on SLP, please contact me directly (Sen Fang). If you have funding support or want to sponsor our laboratory, you can send an email to my advisor and copy me as well. The code will be commercialized and will not be made public.

Demo Overview


ASL
Input: Here's a little example here I'm going to give you some insight on though.
Translation: Here's a little example here I'm going to give you some insight on though.
DGS
Input: Im Südostraum ist es vorteilhaft, ein Hochdruckgebiet kommt von Belgien.
Translation: In the southeastern region it is advantageous, a high-pressure system is coming from Belgium.
KSL
Input: 염려하다.
Translation: worried about.
DSGS
Input: Bei dieser Firma bestelle ich immer die Plastikbecher.
Translation: I always order plastic cups from this company.
LSF-CH
Input: Le leader du parti socialiste.
Translation: The leader of the Socialist Party.
LIS-CH
Input: Io voglio comprare una nuova borsa.
Translation: I want to buy a new bag.
LSA
Input: desayuno.
Translation: breakfast.
TSL
Input: Animasyon atölyesinde 9-12 yaş arası işitme engelli çocuklar animasyon sanatını öğrenecekler.
Translation: Children who are deaf or hard of hearing (aged 9-12) will learn about animation art at the animation workshop.
QP: Synthetic sign language video obtained after processing using style transfer model for reference.

Eight Languages


ASL Sign
Reference Image
ASL
Input: Here's a little example here I'm going to give you some insight on though.
Translation: Here's a little example here I'm going to give you some insight on though.
ASL
Input: Alright our next pose for reviving and waking up in the morning is a seated side stretch.
Translation: Alright our next pose for reviving and waking up in the morning is a seated side stretch.
ASL
Input: People are training you to give you the tools mentally to do the job that you're going to do.
Translation: People are training you to give you the tools mentally to do the job that you're going to do.
ASL
Input: So from here, once I get this out, I rotate the hand just like this.
Translation: So from here, once I get this out, I rotate the hand just like this.
GSL Sign
Reference Image
GSL
Input: Im Südostraum ist es vorteilhaft, ein Hochdruckgebiet kommt von Belgien.
Translation: In the southeastern region it is advantageous, a high-pressure system is coming from Belgium.
GSL
Input: GUTEN ABEND LIEBE ZUSCHAUER, am Anfang Dezember regnet es heute, die Leute gehen nach Hause, das ist es.
Translation: Good evening dear viewers, it started raining in early December and today people are heading home, that's all.
GSL
Input: Im Nordwestraum und später im Südwestraum sowie im Osten wird es etwas trockener.
Translation: The northwest, and later the Southwest, and the east will become drier.
GSL
Input: In der Nacht regnet es besonders im Nordraum und Nordwestraum, im südlichen Teil ist es teilweise klar mit Wolken und Nebel.
Translation: Overnight rain, especially in northern and northwestern areas, partly sunny with clouds and fog in southern areas.
ASL Sign
Reference Image
KSL
Input: 염려하다.
Translation: worried about.
KSL
Input: 한 시간.
Translation: one hour.
KSL
Input: 휴대폰.
Translation: cell phone.
KSL
Input: 좋아, 좋은.
Translation: good, nice.
ASL Sign
Reference Image
DSGS
Input: Bei dieser Firma bestelle ich immer die Plastikbecher.
Translation: I always order plastic cups from this company.
DSGS
Input: Eine Trauung muss vom Zivilstandsamt vollzogen werden.
Translation: The wedding must be performed by a civil registry office.
DSGS
Input: Die Indianer leben in einer anderen Kultur.
Translation: The Indians lived in a different culture.
DSGS
Input: Letztes Jahr erhielt ich als Rendite 500 Franken.
Translation: I made a profit of 500 francs last year.
ASL Sign
Reference Image
LSF-CH
Input: Le leader du parti socialiste.
Translation: Leader of the Socialist party.
LSF-CH
Input: On parle quatre langues en Suisse: français, suisse allemand, italien et romanche.
Translation: Four languages are spoken in Switzerland: French, Swiss German, Italian and Romansh.
LSF-CH
Input: Le chat a des canines très pointues.
Translation: Cats have very sharp canine teeth.
LSF-CH
Input: La fille aime jouer à habiller la poupée.
Translation: The girl likes to play the game of dressing dolls.
ASL Sign
Reference Image
LIS-CH
Input: Io voglio comprare una nuova borsa.
Translation: I want to buy a new bag.
LIS-CH
Input: Ho chiesto a quel sordo che fa il pittore, se puo' farmi in quadro dai toni arancioni.
Translation: I asked the deaf painter if he could paint me a picture in an orange tone.
LIS-CH
Input: Partecipo ad un corso d'inglese di livello avanzato.
Translation: I took an advanced English course.
LIS-CH
Input: Andiamo a camminare in montagna.
Translation: Let's go for a walk in the mountains.
ASL Sign
Reference Image
LSA
Input: photo.
Translation: photo.
LSA
Translation: mock.
Translation: mock.
LSA
Translation: birthday.
Translation: birthday.
LSA
Translation: breakfast.
Translation: breakfast.
ASL Sign
Reference Image
TSL
Input: Animasyon atölyesinde 9-12 yaş arası işitme engelli çocuklar animasyon sanatını öğrenecekler.
Translation: In the Animation workshop, hearing impaired children ages 9 to 12 will learn about the art of animation.
TSL
Input: Ekrandaki engelleri ortadan kaldırdık.
Translation: We eliminated the barriers on the screen.
TSL
Input: İşitme engelli çocuklarımız, başarılarını dünyaya kanıtladılar.
Translation: Our children with hearing impairments proved their success to the world.
TSL
Input: Yedinci sırada yer aldık, ama daha da önemlisi bu, tüm insanlığa engelin bedende olmadığını ilan etti.
Translation: We ranked seventh, but more importantly, it proclaimed to all humanity that disability is not in the body.

BibTeX

@misc{fang2025stablesignerhierarchicalsign,
      title={Stable Signer: Hierarchical Sign Language Generative Model},
      author={Sen Fang and Yalin Feng and Hongbin Zhong and Yanxin Zhang and Dimitris N. Metaxas},
      year={2025},
      eprint={2512.04048},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.04048},
}

@InProceedings{Fang_2025_ICCV,
    author    = {Fang, Sen and Chen, Chen and Wang, Lei and Zheng, Ce and Sui, Chunyu and Tian, Yapeng},
    title     = {SignLLM: Sign Language Production Large Language Models},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2025},
    pages     = {6622-6634}
}