Diffusion weight editing doesn't have to trust CLIP alone: Astra's one-shot programs beat a shared-panel baseline by 10 points on fresh-scene joint success
Synopsis
Having a language model write weight-update programs for diffusion concept editing has been hard to judge, because protection prompts and a numerical fitter sit alongside the program. On 24 TIMED test edits for Stable Diffusion 1.4, one-shot programs written by Astra raised fresh-scene joint success, which requires the subject present and the whole image matching the target description, by 10.0 points over a shared-panel hand-written baseline, while CLIP relative match showed no detectable difference. After probing and rewriting, CLIP rose by 6.2 points but joint success fell by 9.7 points, so program construction and score-guided revision should be evaluated as separate stages.
Interpretation
A complete weight-update program written by a language model can add task-level value in diffusion concept editing, not just a better CLIP proxy score. Concept editing already had closed-form editors such as TIME and UCE with public benchmarks, but a weight-space programming pipeline also contains non-program components such as protection prompts and a numerical fitter, so attribution was unclear. On 24 TIMED test edits, Astra's one-shot programs reached 37.7% fresh-scene joint success against 27.7% for the U1b hand-written baseline that shares the same protection panel and fitter, a difference of +10.0 points with a 95% interval of [+4.7, +15.4]; the same contrast on CLIP relative match was -0.6 points, interval [-7.3, +6.4].
The program organises adjustable directions into components that share one linear invariant and hands them to the fitter to weight independently. For the edit from "a cat" to "a green cat", the Astra program writes 15 further source-target context pairs and 30 preservation texts, and uses TIME's token alignment to form old-state inputs and target-minus-old residuals. It builds an orthonormal basis of the empty-prompt states and the unchanged causal-prefix states, projects its inputs and preservation samples off that span, solves two residual-transfer operators from one shared ridge-regularised covariance, and keeps key and value matrices apart to give four ranked directions; coding all 96 final programs from their syntax trees shows all 24 one-shot Astra programs build an SVD basis and project off it and all 24 encode the empty prompt among protected states, while none of the 24 one-shot Opus programs does either.
Perspective
The conclusions rest on Stable Diffusion 1.4, 24 TIMED test edits, one primary VLM judge (plus five further evaluators on a fixed 144-image subsample), one program per edit and model, and a single rewrite round; the authors explicitly do not compare against published null-space editors such as SPEED or ACE, so what is characterised is construction and revision behaviour in this setting.
The fresh-scene joint-success gain of Astra's one-shot programs is labelled exploratory: Benjamini–Hochberg adjustment across all 288 contrasts gives p = 0.007, Holm within its endpoint gives 0.051 at the boundary, and it does not survive Holm over all cells (0.343). The rewrite stage is not separated by design from the CLIP-guided fitting that follows it, and the complete program's gain is not attributed to an individual component; on the 144-image audit sample, intervals for most evaluators include zero when the attribute is required to fall on the subject itself.
