[신규 모델] SupraLabs에서 Any2Any 모델 패밀리를 시작했습니다!

Repository Files

Installation

bash pip install torch transformers huggingface_hub safetensors Pillow numpy

🧪 Usage Modes

Text generation

bash python run_supra_a2a.py --mode text --prompt "<TEXT>Once upon a time"

Chat mode

bash python run_supra_a2a.py --mode chat

Image reconstruction

bash python run_supra_a2a.py --mode reconstruct --image input.png --out output.png

Text-to-image

bash python run_supra_a2a.py --mode text2image   --prompt "<TEXT>a red square</TEXT><IMAGE>"   --out output.png

Key Insight

This model does not switch between modalities.

It simply:

Predicts the next token.

That token might be: - a word - a visual code - a frame element

Everything is treated equally.

Important Caveats

Attention heads (inferred)

• Default assumption: 4 heads
• May be incorrect depending on checkpoint
• Incorrect value can silently degrade performance

VQ-VAE output activation

Default assumption: - sigmoid (0–1 range)

Alternative: - tanh (-1 to 1 range)

Limitations

• ~30M parameters (small scale)
• 384 token context window
• Low-resolution, abstract image generation
• No RLHF or instruction tuning
• Experimental research prototype

Interpretation

This architecture explores a radical simplification:

Instead of separate systems for vision and language:

👉 everything becomes tokens

👉 everything is modeled by one Transformer

👉 modality boundaries disappear

🧠 Final Take

This is not a production-grade model.

But it is a clean conceptual experiment showing that:

• images can be token sequences
• video can be token sequences
• multimodal learning can be pure language modeling

Feedback welcome!

진심으로 이런 계정들 차단 못 함? 자기 홍보 규칙 없나? 이틀이 멀다 하고 새 모델을 내놓는데 퀄리티도 별로임.

모델이 좋고 쓸만하면 몰라도, 텍스트-이미지 예시는 진짜 웃음 나올 정도로 구림. 2017년 이미지 생성이 더 나았을 듯.