gatilin/mobile_o_0_5b_visual_components

Mobile-O-0.5B trained visual components

Lossless selected tensors from Amshaker/Mobile-O-0.5B: 1,582 entries, 2,070,210,224 payload bytes. Original names/dtypes/values are retained. The full Qwen language model and legacy base_model.model.* copies are excluded.

Includes the trained FastViTHD/MobileCLIP encoder, multimodal MLP, Mobile Conditioning Projector (MCP), complete 28-layer SANA flow transformer and full DC-AE RGB encoder/decoder. This model combines separate understanding and generation branches. The image-understanding encoder is FastViT, and the image-generation codec is DC-AE; extracting only a VAE would omit the trained understanding path and MCP.

The pinned official mobileo/model/builder.py loads mobileoForInferenceLM, whose constructor stores active modules under model.*. base_model.model.* does not belong to that class's active state dictionary. code_provenance.json records the audited loader files. The stale README download filter final_merged_model_23620/* is not used; selected weights come from the pinned repository's actual root safetensors checkpoint. model.safetensors includes the active BatchNorm buffers and is strictly loaded in full.

Usage

Install requirements.txt, then:

from component import load_model
m = load_model('.')
# Official vision preprocessing: pad/resize to 1024px, rescale pixels to [0,1].
features = m.image_features(pixels)
# Official Qwen hidden-state tuple; MCP fuses the final four layers.
conditions = m.condition(hidden_states)
# Generation codec: RGB in [-1,1].
latent = m.encode_latent(rgb)
reconstruction = m.decode_latent(latent)

DC-AE latents have 32 channels at 1/32 spatial resolution. Exact trained SANA latent scale: 0.41407. The model convenience methods apply/remove this scale. Config-only construction prevents downloading base SANA/DC-AE weights: all selected trained weights come from Mobile-O itself. Complete text-conditioned generation still needs the excluded Qwen model and original prompt/scheduler pipeline.

validate.py checks all source tensor payload hashes, strict native loading, native-resolution FastViT, MCP layer selection/weighted fusion, square/rectangular DC-AE roundtrips and complete SANA predictions. validation.json records actual results and test limits; these smoke checks are not generation-quality benchmarks.

License and attribution

The HF source card declares CC BY-NC 4.0 for weights (LICENSE, preserved SOURCE_MODEL_CARD.md). The official Mobile-O code repository separately carries CC BY-NC-SA 4.0 (MOBILE_O_CODE_LICENSE); bundled original code retains those terms. The FastVLM-derived understanding tower also retains the Apple ML Research Model License, which limits model derivatives to noncommercial scientific research and academic development. APPLE_FASTVLM_MODEL_LICENSE, APPLE_FASTVLM_LICENSE, APPLE_FASTVLM_ACKNOWLEDGEMENTS, and original source headers preserve those upstream terms and notices. See LICENSES.md and NOTICE_MODEL for component scopes and extraction changes. Source-card CC BY-NC metadata does not replace inherited research restrictions. Original authors: the Mobile-O team and credited upstream projects. Native constructor code and trained tensors are unchanged; packaging/loader glue and validation scripts are added here.