RFD 1136: Views are named in the ecosystem’s camera vocabulary
Problem
This workspace named views in degrees. “Pitch -90, yaw 0” says nothing about what a picture shows, and a reader has to hold the up axis and the rotation order in their head to find out.
The cost was not cosmetic. Loop 1 scored view 0 of its sequence for three rounds and got 0.0 each time. View 0 is a camera directly beneath a standing figure. The degrees were on the page the whole time.
Decision
Name a view in the vocabulary fal’s Multiple-Angles LoRA is conditioned on: 8 azimuths at 45 degree steps, 4 elevations at -30, 0, 30 and 60, and 3 distances at 0.6, 1.0 and 1.8. It then reads the same to a person, a tagger and a generator.
camera_prompt converts parameters to the prompt and camera_parameters converts back, both exactly. A camera the table does not name is a refusal rather than the nearest phrase, because a snapped name describes a picture nobody rendered. describe_camera is the lossy direction and says so, naming the band that holds a camera and refusing outside it. sample_camera draws inside a band, and its seed is required.
The 96 cells are enumerated in test_weft_loop.py, not sampled, and that found the front view’s band wrapping through zero.
Only the phrase table is used. Qwen-Image-Edit is blocklisted and those weights are gated, so neither is loaded. logbook/CITATION.cff cites it.