InternVL-U Generation DiT — Native Component

Component validation: 2026-10-28.

This repository exports the complete, trained InternVL-U generation decoder. It complements the separately released vision bridge and Qwen-Image VAE. It accepts native VLM conditioning and VAE latents; it is a component rather than a standalone text-to-image pipeline.

Included scope

• All 20 native double-stream transformer layers, input/time/condition/output modules, and the learned condition-padding token.
• 1,808,034,880 parameters in 654 tensors; original mixed BF16/F32 storage is preserved. The complete original file matches the pinned official LFS SHA-256.
• Unmodified native computational sources, the native DPM multistep flow scheduler configuration, and the original guidance pipeline source.

The memory-efficient loader reconstructs the native, deterministic RoPE frequency cache after meta initialization. This cache is not a learned checkpoint tensor; the weight file stays byte-identical to the official source.

Native interface and dependencies

Inputs are [B,16,H,W] VAE latents with even spatial dimensions, [B,L,4096] conditioning features from the matching InternVL-U language model, and native timesteps. Semantic image tokens require their image grids and masks. Reference VAE inputs retain the upstream geometry constraints.

The VLM, vision tower, processor/tokenizer and VAE weights are excluded. Preserve the original conditioning-layer selection, Qwen-Image VAE normalization, DPM flow-sigma settings and three-branch guidance. The guidance pipeline source includes high-timestep normalization; replacing it with generic two-branch CFG changes inference.

import sys, torch
from huggingface_hub import snapshot_download
path = snapshot_download("gatilin/internvlu_image_dit")
sys.path.insert(0, path)
from component import load_component
dit = load_component(path, dtype=torch.float32, device="cpu")
# Interface smoke fixture only: random features are not a real text prompt.
latents = torch.randn(1, 16, 4, 6)
features = torch.randn(1, 4, 4096)
prediction = dit.predict(latents, 500.0, features)
scheduler = dit.make_scheduler()

For full inference, use the matching upstream pipeline and real VLM conditions. The helper preserves a narrower prediction interface; the original native pipeline is included for guidance and sampling integration.

Source and reproducibility

• Official source: InternVL-U/InternVL-U.
• Weight revision: f012d760e69712bb47f7d3d09a24280f346cee01.
• Source file: generation_decoder/model.safetensors; official full-file SHA-256: 3d4d40703e8a8d805ff41a75baa17085fc54eaec7f3a8658b24d80578489cb3b.
• Export SHA-256: 3d4d40703e8a8d805ff41a75baa17085fc54eaec7f3a8658b24d80578489cb3b.

See extraction.json, source_code_manifest.json and requirements.txt. Loading requires explicitly importing the provided component code. No complete language model is bundled.

Validation and limits

All 17 recorded checks passed: strict native loading, source-weight integrity, full-depth prediction, independent attention comparison, condition/time responses, batch isolation and native sampling trajectories.

Tests use real checkpoint parameters, CPU float32 execution and small synthetic conditions/latents. They do not measure real-prompt generation, waveform/image decoding, perceptual quality or complete upstream task performance. Storage dtypes remain unchanged. See validation.json; rerun with python validation_reference.py internvlu.

Attribution and licenses

Native InternVL-U source and weight attribution is retained under the upstream MIT license. Qwen-Image, PixArt-Sigma and Hugging Face-derived source files retain their Apache-2.0 notices; see LICENSE-APACHE-2.0 and file headers.