RFD 1117: Text-to-image chains into Image-to-3D, one URL resolver

3DAIGC-API’s DINOv3 adapter

Problem

A completed krea2_turbo_text_to_image row offers “Use for Image to 3D,” chaining straight into trellis2_image_to_textured_mesh. Two regressions kept recurring: a browser fetch reading a filesystem path (outputs/images/...) instead of the job’s own download URL, and a duplicate DINOv3 feature-extractor class drifting from the shared one.

Decision

resolveTextToImageDownloadUrl(task) is the one URL resolver, for preview, download, and chain alike. No code fetches an outputs/images/... path directly. The chain button fetches through that resolver, sets the target task type to image-to-3d, attaches the fetched image as a File, expands the Tasks panel, and scrolls to the new-task form. A text-to-image completion never auto-loads into the 3D viewport on its own. The T-pose and A-pose chips stay mutually exclusive.

On the backend, get_dinov3_encoder_layers() is the only path to the DINOv3 layers; image_feature_extractor.py holds no second extractor class. YAML init_params metadata passes through _filter_init_params before reaching an adapter’s __init__. A pip or TRELLIS.2 edit runs verify_hf_conditioning.py, then restart_services.sh, before merge.

Related

RFD 1038 gives the TRELLIS.2 model image this chain’s target task runs on. This RFD replaces krea2-text-to-3d-pipeline-protected.mdc, listed in RFD 1112.