Tuna-2 (GitHub Repository)

Tuna-2 is part of the Tuna series aimed at multimodal understanding and generation, and it simplifies the architecture by progressively removing stages of visual encoding.

• Removing the VAE produced Tuna-R, which performs unified multimodal processing in pixel space using only a representation encoder.
• Tuna-2 goes further by bypassing even the representation encoder, applying patch embedding directly to raw images.
• As a result, Tuna-2 outperformed both Tuna-R and the original Tuna across various multimodal benchmarks.

Inference is handled through a single script, scripts/launch/predict.sh, which supports the t2i and edit tasks. Variants are divided into none_encoder (Tuna-2), siglip_pixel (Tuna-R), and vae (Tuna), with 7B and 2B sizes available. However, 2B is only available for the vae variant.

Resolution support includes combinations at the 512 and 1024 levels, offering 512x512, 448x576, 576x448, 384x672, 672x384 and their corresponding 1024x1024, 896x1152, 1152x896, 768x1344, 1344x768 configurations.

The video generation model has not been released due to organizational policy, but the training and inference code is provided. Instead of the full production-trained weights, a foundation checkpoint is released first, in which some layers of the LLM backbone and diffusion head (flow head) are removed while the remaining components such as the vision encoder, projection, and embedding are kept intact, designed so that it can be quickly restored with short fine-tuning on one's own data. Plans call for releasing the fully restored weights at a later date.