Three live-run findings from the rental:
1. Block steering said "profile not ready (needs 2+ rated runs)" with 16
paired runs. The v1 liked/disliked-EMA attribution had no contrast when
ratings cluster on one pole (pole-vs-all-runs-mean cancels to zero) and
dropped rewards in (0, 0.3) entirely. Attribution is now the per-block
COVARIANCE between reward and activity over a rolling 32-run history:
every admissible reward contributes with its actual value, contrast
comes from reward variance, and the sampler log now states the real
blocker ("all rated alike - rate both good AND bad runs") instead of a
misleading run count. v1 profiles are discarded on first pairing.
2. output_guidance stayed "not ready (needs 10+)" forever on LTX-AV: the
end-of-run x0 snapshot was gated on output samples being a plain
torch.Tensor, but AV chain output is a NESTED tensor (video+audio) - so
the output value function never received a single sample. The snapshot
now extracts the video stream from nested outputs.
3. FunPackNVFP4ModelLoader hit a size-mismatch wall (9-vs-6 ada params,
connector dims): it loaded the state dict without the safetensors
METADATA that carries the LTX-AV architecture config, so detection
built the wrong model variant. Metadata now flows through, mirroring
comfy.sd.load_diffusion_model.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
RLHF at the model level: which REGIONS of the transformer were doing the
work when a generation came out good or bad, and steer accordingly.
1. Measure: tagged forward hooks record each block's relative video-residual
delta per model call (near-zero cost; covered by the existing leak-strip).
2. Attribute: the end-of-run fingerprint is staged (block_snapshot.pt, same
stage-then-pair convention as x0_snapshot) and paired with the rating
reward on the next Refiner cycle into liked/disliked/all-runs EMA
profiles; when output_guidance's value function is trained, an innermost
model-function wrapper additionally credits blocks per step by
(value delta) x (centered activity) - converges in a handful of runs.
3. Steer: vx' = vx_in + g_b*(vx_out - vx_in) on the video stream only,
audio untouched; scores zero-mean/unit-max so emphasis is redistributed
rather than inflated, gains hard-capped at +/-10%. Exact no-op until the
profile has 2+ paired runs; blocks at gain 1.0 get no hook at all.
Chain Sampler toggle + strength (append-only widgets), exposed same-change
in the editor's Chain Sampler Guidance section via the sampler_inputs
bridge. New sidecars ride the defensive <key>.* sweep in
delete_refinement_key, so atomic key deletion covers them.
Tests: toy AV-block model driving the real hooks (recorder ranking, steer
video-only scaling + native path for unit gains, credit wrapper, snapshot/
profile/scores/gains roundtrip, mild-positive band, depth-mismatch reset,
leak-strip tagging). Full suite failure set identical to the pre-existing
baseline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
