gatilin/openvision3_base_latent_visual_encoder

OpenVision 3 Base latent visual encoder

All 150 trained entries from the pinned official checkpoint, exported as safetensors with original names, shapes, dtypes and logical values. Noncontiguous source storage is packed contiguously without transposing values. The complete original PyTorch file matches its official LFS SHA-256; every logical tensor digest is recorded in tensor_sha256.json.

This checkpoint is a latent ViT encoder: 12 layers, width 768, 16 input channels, a 32x32 latent grid, 2x2 patches, 256 output tokens and a pooled 768-dimensional projection. It contains no RGB VAE, reconstruction decoder, diffusion model or language model. Its representation is intended for both understanding and generation; this package alone is not a complete generation pipeline.

Usage

Install requirements.txt and download the repository files:

import torch
from component import load_model
model = load_model('.')
latents = torch.randn(1, 16, 32, 32)
with torch.inference_mode():
    pooled, tokens = model(latents)
print(pooled.shape, tokens.shape)  # (1, 768), (1, 256, 768)

The official RGB example first applies an external FLUX VAE, using its posterior mean multiplied by its scaling factor. That pinned example does not apply a latent shift. External VAE and matching image-reconstruction decoder weights are not included.

Verification

validation.json records nine passed CPU float32 checks covering original LFS checksum, all 150 values/dtypes, strict native loading, input/output geometry, finite embeddings, exact native pooling/projection and batch independence. The bundled model is the exact official custom OpenVision VisionTransformer, not a substituted RGB OpenCLIP model.

Run python verify_release.py to verify every original tensor hash, strict loading, full encoder inference, pooling and batch independence. recovery_verification.json records the same checks on this publication copy after local cache recovery. Tests use synthetic latent inputs and do not establish RGB reconstruction or generation quality.

Source and license metadata

Model attribution: UCSC-VLAA/OpenVision 3 authors; original repository and revision linked above. The inspected upstream checkpoint repository does not declare a weight license. This export preserves that status and asserts no additional weight-license grant (LICENSE). Official project code is Apache 2.0 (CODE_LICENSE); original headers and code_provenance.json retain attribution. The code license is not presented as a new declaration of the checkpoint weight license.

Inherited OpenCLIP and Meta MAE source notices are additionally retained in OPENCLIP_LICENSE and MAE_SOURCE_LICENSE; the original module headers remain intact.