exp010_controller — THE STEP-GATED DIFFUSION CONTROLLER + generation-side
band-lesion battery (CANDIDATE, 1-seed battery)
The artifact. StepGatedSampler — DDIM sampling where EVERY denoising
step sets the band crossfade windows from the current timestep, activating
the trained band experts coarse-to-fine across the trajectory. The stage's
controller, shipped as a usable module.
The question. exp009 showed per-step eps-MSE is nearly blind to band behavior (0.05–0.2% moves). What do the bands do IN IMAGE SPACE? Battery: exp008's 2-seed-confirmed mb3 stack; arms all_on / lesion each band / frozen; paired seeds; judged by round-trip grounding, lesion-vs-all_on LP/HP decomposition, and within-prompt diversity (6 prompts × 4 seeds).
Results (results.json):
| arm | grounding gap | diversity | LP-change vs all_on | HP-change |
|---|---|---|---|---|
| all_on | +0.2135 | 0.196 | — | — |
| lesion band0 (LOW noise) | +0.1979 | 0.208 | 0.0036 | 0.0031 |
| lesion band1 (MID) | +0.1821 | 0.217 | 0.0184 | 0.0039 |
| lesion band2 (HIGH noise) | +0.1782 | 0.297 | 0.0727 | 0.0051 |
| frozen | +0.1247 | 0.332 | 0.0804 | 0.0052 |
Findings.
- The controller lifts grounding +0.089 over frozen (+0.1247 → +0.2135) and every band lesion cuts it monotonically — all three bands carry generation-relevant signal that eps-MSE could not see. The image-space instrument was the right upgrade.
- Coarse-to-fine confirmed by lesion signature ratios: the HIGH-band lesion is 14× LP-dominated (0.0727 LP vs 0.0051 HP — structure/layout), the LOW-band lesion is ~1:1 detail-leaning (0.0036/0.0031), MID intermediate. (Raw prereg G2b "LOW moves HP most in absolute terms" missed — layout shift mechanically drags detail with it; the per-lesion ratio is the honest read, disclosed.)
- Diversity interpretation flip (G3 miss, informative): the eps-trained HIGH band CONCENTRATES generations toward the conditioning (its lesion raises diversity 0.196 → 0.297; frozen is most "diverse" because least grounded). A diversity-CARRYING high-noise expert remains a training- objective goal (nothing here trained for diversity), not a refuted property — the docket's diversity/blob supervision is the open path.
Caveats. 1-seed battery (one noise-seed bank, n=24 + 6×4 diversity probe); one checkpoint (uniform-trained mb3_s0; the roles checkpoint runs the same battery as an amendment); CLIP-L gauges. Cost ≈ 0.2 GPU-h.
AMENDMENT: the ROLES checkpoint through the same battery (results_roles_s0.json)
Image-space lesion signatures are IDENTICAL to the uniform checkpoint's (band2 LP-change 0.0724 vs 0.0727; all bands within a few %), and role training slightly COST grounding (all_on +0.1996 vs uniform +0.2135). Frequency-reweighted role pressure is now confirmed inert on BOTH instruments (per-step gauges, exp009; generation-side battery, here). The role thesis's open path is qualitatively different supervision (segmentation blobbing / diversity pressure — the stage's exp011+).