UniLIP-1B trained visual components

Lossless extraction from kanashi6/UniLIP-1B. All 1,190 selected entries retain their original names, dtypes and tensor bytes (2,312,193,222 payload bytes).

Includes the trained InternVision tower, space-to-depth multimodal projector, learned 256 queries, six-layer Qwen connector, caption projector, custom DC-AE decoder with three residual adapters, and the complete 28-layer SANA flow transformer. The main Qwen language model and LM head are excluded. This is a visual-component package; complete text-conditioned generation requires the official language model, preprocessing and sampling pipeline.

The decoder is not a standalone RGB encoder VAE: image encoding follows InternVision → patch merge → multimodal projector → custom residual/down-projection, yielding 32-channel DC-AE latents. Input 448×448 gives 16×16 latents and 512×512 decoded RGB. Flow-space latents use unilip_factor = 10.6; divide by 10.6 before decoding generated SANA latents.

Loading

Install requirements.txt. The official UniLIP repository has no root code license, so its custom decoder source is intentionally not redistributed here. Obtain the official source yourself at commit 937ee73991d68e1a53c1ac156e1a62e6b251e6aa and pass the local file explicitly. The loader checks its SHA-256 against external_decoder_source.json. No constructor downloads extra weights.

from component import load_model
model = load_model('.', '/path/to/UniLIP/unilip/model/vae_modules.py')
# pixels: normalized BCHW, square dimensions divisible by 28
latents = model.encode(pixels)
rgb = model.decode(latents)
# hidden_states: supplied by the original language model
conditions = model.condition(hidden_states, attention_mask)

validate.py uses UNILIP_DECODER_SOURCE=/path/to/vae_modules.py. See validation.json for exact tested shapes and limitations, and the extraction/mapping/hash files for byte preservation. Tests cover strict loading, full-size vision and decoder inference, original patch ordering, bidirectional connector, learned queries, and a complete SANA prediction. They do not establish downstream generation quality.

Weights follow the source model card's Apache 2.0 declaration; SOURCE_MODEL_CARD.md preserves it. Bundled InternVision source carries its original MIT headers and INTERNVL_LICENSE. Other native constructors are supplied by Apache-licensed Transformers/Diffusers. The external UniLIP source is excluded; this package does not grant rights to that separately obtained code.