DeepGen-1.0 SD3 DiT
Lossless extraction of SD3-Kontext 24-layer DiT + SCB connector + learned meta queries from deepgenteam/DeepGen-1.0-diffusers, pinned revision 85c2ed257b9406d8531965c68e140dc7fdb81382.
This package is a trained generation component. It requires original language hidden states and a compatible external codec; language/vision/codec weights are excluded.
Included and precision
The complete trained generation blocks, patch input/output, time embeddings, normalization/modulation, native positional logic and conditioning modules are preserved. Original tensor names, dtypes and values are unchanged. extraction.json and tensor hashes record provenance. component_spec.json defines the exact conditioning and latent contract.
Load offline
from component import load_model
model = load_model('/path/to/this/package') # preserves checkpoint precision
# For CPU float32 numerical work, explicitly pass dtype=torch.float32.
# seq, pooled = model.condition(original_vlm_hidden_states)
# prediction = model.predict(latents, timestep, seq, pooled_projections=pooled)
predict returns the native prediction, before CFG and scheduler stepping. For Yak, supply packed latents and img_ids/txt_ids; use get_schedule with the original image token count. For other models, make_scheduler() reconstructs the stored original scheduler, including flow shifts and prediction conventions. DeepGen reference inputs use its native per-sample reference lists; UniPic2 uses its native BCHW reference tensor. SANA packages expose no reference-latent editing API.
Latents and conditioning
Concatenate the six official SCB VLM hidden-state tensors along the feature dimension, preserving their original descending-layer order. Input is hidden states, not text token IDs.
{
"channels": 16,
"scaling_factor": 1.5305,
"shift_factor": 0.0609,
"encode": "(codec.encode(rgb) - shift_factor) * scaling_factor",
"decode": "codec.decode(latent / scaling_factor + shift_factor)",
"codec_weights_included": false
}
Do not replace this codec or its scaling/shift/statistics by an arbitrary VAE. Cross-attention hidden states must come from the original trained language-conditioning path. This component alone does not accept text prompts or retain general multimodal understanding.
Validation and limitations
See validation.json and verify_release.py for the checks actually run. Validation uses the real, complete trained weights and native model depth, with small synthetic latent/condition fixtures. It separately checks original payloads, strict loading, conditioning, native attention versus explicit softmax, CFG and native scheduler behavior where applicable. It is not a language-conditioned full-resolution generation test, GPU image-quality benchmark, or paper-score reproduction. VAE/codec decode is covered by the referenced codec package, not re-claimed as tested here.
Source and notices
HF checkpoint license metadata and original code/weight notices are retained. Mobile-O inherited Apple model terms remain included. DiT training ancestry does not erase upstream model terms. For BLIP3o-NEXT, official code has no separate license file; source URLs/hashes are provided without redistributing those files, and the loader constructs the recorded architecture using diffusers primitives.