Adversarial review of the leak fix found the new tests only exercised REINS
stacked on itself, leaving 6 of 7 install mechanisms and the entire
optimized_attention_override chain (av_decouple/attn_temperature/q_steering)
untested. Add a test stacking REINS -> av_decouple -> q_steering -- different
mechanisms chaining onto each other, the realistic multi-run leak shape -- and
confirm one strip call unwinds both the per-block dit dict and the single-slot
attention override chain.
patches_replace["dit"] and optimized_attention_override installs had no tag+strip
protection, unlike the scene-wrapper and block-hook mechanisms fixed for the same bug
class earlier. An interrupted run left these hooks on the shared model; the next run
chained a new layer onto the leaked one instead of starting clean, growing unbounded
across interrupted runs -- the likely cause of the live "potential memory leak...
circular references" warnings and slow full-GC pauses.
REINS ("h3_repr_steering") only ever captured while its own toggle was on --
turning it off also stopped it from learning, with no way to bank rating
history separately. New h3_repr_steering_passive_capture keeps the existing
unconditional per-block capture running (forcing strength to 0 so nothing is
ever injected) even while the main toggle is off. Exposed a real dormant bug
in REINS' own console print: it claimed "applying learned direction(s)"
whenever a direction existed, regardless of strength -- harmless before
because the toggle and strength widget always moved together, not once
passive capture legitimately runs at strength 0 with real data available.
Fixed to report "capturing only" in that case, mirroring h3_q_steering's own
already-fixed version of the same print.
Also: h3_q_steer_block/h3_q_steer_strength (from the H3 query steering
feature) were never wired into the Movie Editor's Settings UI, only exposed
as raw ComfyUI node widgets -- missing this project's own rule that every
new sampler toggle lands in the Editor in the same change. Added both,
plus the new passive-capture checkbox, to Engine Settings > Chain Guidance
alongside REINS' existing controls.
Two real faults found by a third adversarial review round on 4154746:
- _install_h3_av_decouple's own bail check (`_strength <= 0.0`) disagreed with
_sample_chunk's install guard (`h3_av_decouple > 0.0`) for a NaN strength:
every ordered comparison is False for NaN under IEEE 754, so `<= 0.0` let a
NaN through as "on" while the guard treated the same value as "off" --
reproducing the exact av_decouple-defeats-q_steering bug the last fix
closed, just reached through a different value. Both now use the same
`not (x > 0.0)` predicate so they can't disagree.
- h3_q_steering was installed before h3_explore_temperature, so if both
target the same block, temperature's Q-scaling reached q_steering's
capture before it ran -- banking a temperature-distorted "natural Q"
descriptor, contradicting the mechanism's own documented ordering promise
(capture reflects the network's natural Q, same claim h3_repr_steering
makes for hidden state). q_steering is now installed last (outermost),
after both av_decouple and attn_temperature, so it always sees Q first.
av_decouple still can't compose with q_steering at all regardless of
order (unrelated masking issue, already handled by skipping install).
Regression tests for both, plus a NaN case added to h3_av_decouple's own
existing zero-strength no-op test.
Two real faults found by a second adversarial review round on 608a97a:
- h3_av_decouple, installed after q_steering, becomes the OUTER
optimized_attention_override and always calls its inner override with a
non-None mask for the video-row partition -- exactly the case q_steering's
own override treats as "already handled, back off". With both enabled,
q_steering captured and injected nothing while still printing "applying"
and paying the clone/hook cost. The two can't currently compose, so
q_steering is now skipped entirely (with a declared reason) whenever
h3_av_decouple is also active, rather than silently no-op.
- conditioning.py's q_steer commit call was silent on no_pending/no_key,
justified by a comment claiming REINS' own commit is silent there too --
that was wrong, REINS is loud on exactly those outcomes. Fixed to match.
Also tightened the shape-mismatch warning's wording: it checks total element
count, not the heads/head_dim split, so a same-total-different-split
mismatch would still reshape onto the wrong head boundaries undetected --
documented as a known narrow gap rather than fixed, since a real fix would
mean changing the persistence format REINS itself also depends on.
Added regression tests for the av_decouple interaction (both directions:
guard fires when it should, stays quiet when it shouldn't).
Three real faults found by adversarial review of 299941a:
- sample()'s own default for h3_q_steer_block was "0,1" while INPUT_TYPES
declared "" -- any saved workflow predating this widget omits the key
entirely, so ComfyUI falls back to sample()'s default, silently turning on
capture/clone/hook for every H3 generation with a refinement key. Now both
default to blank, matching every sibling knob in this file.
- The "applying learned direction(s)" console line fired whenever a direction
was available, not when it was actually applied -- strength=0 still printed
it, which is the exact silent-success-message trap this project has already
been burned by. Now split into an "applying" line (strength>0) and a
"capturing only" line (strength=0).
- A direction banked under a different heads*head_dim layout (model swap on
the same refinement_key) hit an unguarded .view() and would hard-crash
generation instead of degrading. Now checked and skipped with a one-time
notice, same pattern as the existing batch>1 guard.
Added regression tests for all three, plus two coverage-gap tests the review
flagged (kind isolation against REAL persistence, not mocks; two named blocks
steering with independent directions in one run).
Same liked-minus-disliked mean-difference as h3_repr_steering, captured off Q inside
optimized_attention_override instead of a block's hidden state, own kind="q_steer" sidecar
so the two vector spaces never collide. Only Q is biased: a uniform additive bias on K is a
provable softmax no-op, and on V it's mathematically identical to h3_repr_steering's own
mechanism, so neither is a genuinely different lever. Off by default (blank block spec).
h3_repr_steering.py's persistence functions take an optional kind= param (default unchanged)
so this reuses capture()/direction()/commit() verbatim instead of a parallel implementation.
optimized_attention_override was a single dict key that h3_av_decouple claimed
outright -- a second mechanism on the same slot would silently replace it with
no error, the exact bug class dit_patches chaining already guards against
elsewhere. Fixed so any install chains through whatever's already there.
New h3_explore_temperature (+ block spec): flattens attention's own softmax at
named blocks instead of adding ungated latent noise -- variety stays bounded
by the prompt since it can't attend to anything not already in context.
Deliberately NOT wired to REINS or any rating data: neither mechanism's
behavior is understood well enough yet to make coupling them debuggable. A
plain manual dial, same as block repeat. Off by default.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
No way to read H3's actual trained attention-logit scale from this machine
(checkpoint isn't here) -- so normalizing the knob to 0-1 against a hidden
made-up multiplier (12.0) was a second guess wearing a calibrated costume,
and exactly the kind of thing that produces another flat live-tested curve.
The knob is now the raw softmax logit penalty directly (0-50, step 0.5), so
a live sweep searches the actual unit instead of my conversion factor.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Live 0-1 strength sweep came back flat -- confirmed the mechanism, not just the
value, was wrong. H3's Q/K get RMSNorm'd before attention, so scaling a key
toward zero pushes its logit toward the *mean* of an already near-zero-centered
distribution, not toward excluded. Switched to what actually suppresses a
softmax entry: an additive negative bias on the targeted logits, still passed
through the same per-query-group split (still no S^2 mask, still same total
FLOPs -- the bias is a per-key vector broadcast over the group's queries).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
H3 has no cross-attn to mask (unlike LTXAV's Bounded Attention), so a dense
attention mask would be GB-scale over its ~37k-token packed sequence. Uses
comfy's own optimized_attention_override extension point instead: runs the
same joint attention as separate video-query/audio-query passes against a K
where the other modality's rows are scaled down, at the same total FLOPs as
one pass. New h3_av_decouple sampler input + Engine Settings entry, off by
default (0 installs nothing).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The sweep's rating feature was wired against REINS' single 'pending'
capture (one slot per key, overwritten by whatever generates next),
which meant only the MOST RECENT result in a whole batch could ever be
rated -- the opposite of what batch training actually does: rate every
result you generated, independently, in any order.
New h3_repr_capture_slot sampler input (blank by default = unchanged
ordinary one-at-a-time flow) saves a run's capture under an explicit
slot instead of the single pending. New h3_repr_steering.save_capture_slot
/ commit_slot / discard_slot mirror save_pending/commit but keyed, so a
whole batch stays independently rateable the same way batch_training.py's
per-item .cond.pt files already do for value-function training.
New server routes commit/discard a slot directly -- no extra generation
needed, unlike the old 'fire another render with rating attached'
workaround that existed only because of the single-slot limitation.
Every gallery card is now rateable (Like/Dislike/Skip), not just the
last one.
h3_repr_steering_block goes from a fixed combo (18 candidate blocks) to
free text -- a single block, a range, or a comma list, same syntax as
Repeat block(s) (31-40 / 4,5,6). Each named block steers with its OWN
learned direction from its OWN liked/disliked history at that block --
there is no shared 'the direction', just N independent ones applied at
their own positions in the same forward pass. A block short on rated
data at itself just captures, same as any unnamed candidate block.
CANDIDATE_BLOCKS widened from 18 curated blocks to all 50 -- every block
is now both capturable and steerable, not just the ones on the old grid.
Fixed a pre-existing sync gap: h3_block_repeat_last_steps was never added
to settings_card.py's ownership/defaults tables, breaking two golden-file
tests as soon as they ran (test_settings_card_noise.py).
Same widget, sign now carries direction: positive N = final N steps
(tail, refine only -- what every laststeps test so far has used).
Negative N = first |N| steps instead -- apply while structure is still
forming, then get out of the way and let the untouched rest of the
schedule resolve it, instead of the perturbed step being close to the
last word on the output.
Widget min widened 0 -> -50 (backward compatible: 0 still means every
step, existing positive values unchanged). Sweep panel accepts negative
laststeps in the config line (e.g. 40-41;noseam;1;-2).
Every step was being repeated, including the earliest near-pure-noise ones
where a block is still deciding layout, not detail -- the likely reason
repeat re-rolls the shot instead of sharpening it. New h3_block_repeat_last_steps
(0 = every step, unchanged default) windows the repeat to the tail via the
exact step index off transformer_options[sample_sigmas/sigmas].
A different operation, not a re-spelling. Per-block repeat hands every
block its OWN doubled output, which is a distribution the NEXT block has
never seen, so a ten-block span creates ten seams. A span loop leaves
every block receiving input from its normal predecessor and creates
exactly ONE, at the wrap from last back to first.
That is the obvious candidate explanation for the sweep's central
puzzle: 31,40 and 35,40 survive two extra passes while 31-40 collapses
at the same count. Same total extra block evaluations, so the same cost
— only the seam count differs.
The span's first block does the work (walking
model.model.diffusion_model.blocks, which the hook cannot reach since
original_block closes over one block only) and the rest of the span
becomes a pass-through, because the model's own loop still calls them.
Refuses a non-contiguous selection with a note rather than silently
looping min..max, which would run the blocks deliberately left out.
Under video-only, text and audio rows get the span ONCE — not zero
times, which would skip those blocks entirely for audio and is a
different, worse intervention than "do not repeat them".
Band 31-40 is the best video result in the sweep and invents unprompted
dialogue on every single run of it — a trade that the bisection proved
cannot be dropped by narrowing, because neither half reproduces either
effect.
H3 packs video, text and audio into ONE sequence, so a plain repeat
doubles the block for the audio rows too. This restores text and audio
rows to their single-pass value after the extra passes, keeping the
repeat for video only.
It does NOT undo audio's influence on the second pass — attention had
already mixed the twice-processed audio rows into the video rows before
they were restored — so it fixes the audio that comes OUT, not the video
path. Whether the video gain survives that is the experiment.
Off by default so the measured sweep stays reproducible, and it falls
back to repeating everything when the mask code meets a shape it has
never seen, rather than silently dropping the feature.
They were sharing one toggle on the argument that a second switch for
the same decision is worse UI. That was wrong for a specific reason: the
two probes cost wildly different amounts — the detail probe copies one
latent per run, the block probe gathers ~17k x 5120 rows per block per
step across 18 blocks — and allocator churn on that scale can change
which GPU kernels get selected, so a probe can move the numbers it is
measuring. Answering "is the measurement causing the variation?" needs
one running while the other is off, which a shared toggle made
impossible to perform.
Also: the noise floor is now measured from your own repeat rows (median
of 1 - structure across pairs where nothing changed) instead of graded
against a fixed 0.98 that no real rig reaches. Other rows report their
movement as a multiple of it, so "structure 0.42" reads as "6x your
noise floor" — which is the part that means something.
The conditioning fingerprint was a formatted string, so it was only ever
equal or not — encoder float wobble read exactly like a new prompt, and
every row claimed the prompt changed on four runs where it demonstrably
had not.
It is numbers now (numel, mean, std per entry), compared with a
tolerance, and the row reports the SIZE of the move. That separates
three cases the old label collapsed into one:
under 0.1% encoder wobble — not reported at all
a few % "conditioning drifted 2.1%" — something rating-driven
moved it (h3_phrase_emphasis lives on the conditioning
and updates as ratings land), worth naming but it does
not invalidate the comparison
token count "prompt/scenes changed" — structurally a different
or entry generation, and only this still voids the A/B
count differs
The influence probe measured the stack as heavily back-loaded — blocks
48 and 49 alone carry about a third of all residual movement, and the
busiest twelve are almost all in the 40s — while REINS defaults to block
25 and the picker stopped at 40, so it could not reach where the model
does most of its work.
CANDIDATE_BLOCKS now covers 40-49 individually (the quiet early half
stays on the coarse grid). Capture has to cover them too: direction()
reads descriptors at the exact block being steered, so a picker entry
outside the capture set would say "not enough data yet" forever. A test
pins that the two sets cannot drift apart — the same silent-substitution
class as the rating combo that discarded every digit.
Node and Editor choice lists verified byte-identical: a mismatch there
makes ComfyUI's widget validation quietly swap in the default.
The run fingerprint captured only scalar arguments, and the prompt
arrives as conditioning tensors — so two runs of completely different
text compared as a repeatability check, and the resulting structure drop
(0.44-0.93) read as "generation is not reproducible here" when one row
in the same batch showed a bit-identical pair (structure 1.000, detail
x1.000, edge +0.000), proving it demonstrably is.
Fingerprint now includes a conditioning stand-in: per-entry shape + mean
+ std, which differs between any two real prompts and costs two
reductions on a tensor already resident. The row says "prompt/scenes
changed" and refuses to call the pair a noise-floor reading.
Live data showed block 1 at 6.01 and block 0 at 1.83 against ~0.2-0.35
everywhere else, with block 1's liked/disliked gap (+5.82) as large as
its whole measurement. That is not "the first two blocks do 30x the
work" — ||f_i||/||x_i|| divides by a residual stream that GROWS with
depth, so early blocks score high because the denominator is barely
formed. Ranking on it just sorts the earliest blocks to the top, and a
high-variance early block fakes a large taste correlation at low n.
The hook now stores both norms instead of their ratio, and profile()
adds `share` — each block's raw ||f_i|| over every block's, one
denominator for the whole stack. The panel ranks and displays that
("2.3% of the movement"). The ratio is kept, since it is still the right
answer to "how much did this block move the stream relative to what the
stream was".
Checked against data shaped like the report: ratio ranks 0,1,2 (all
artifact), share ranks the blocks that actually move more.
Every row read "no repeat → no repeat" because the label only ever
tracked one knob — the block-repeat spec — so a comparison where
anything else changed, or nothing did, described itself as unchanged and
told you nothing about what it compared.
sample() now snapshots its scalar arguments (locals() at the top IS the
argument list, so no name list to go stale), and each row keeps the DIFF
against the previous run. The row titles itself:
h3_block_repeat: — → 40 sharper — +204% detail, on existing edges
nothing changed stable — two identical runs match, so a
real A/B can be trusted
seed 111 → 222 different seed — not comparable
The "nothing changed" case is the valuable one and was previously
unreachable: two identical runs measure how much generation varies on
its own. If that comes back unstable, no A/B on the setup means
anything, and the row says so instead of letting every later comparison
quietly inherit the noise.
A block that writes its residual in place leaves args["img"] already
holding the post-block values, so every delta came out exactly zero —
and a zero delta gives a zero cosine, so magnitude AND novelty both read
0.0000 / +0.00 on every block. That looks like "this model has no block
structure", which is a finding, when it actually meant the probe
measured nothing. The exact silent-no-op class this probe exists to
catch, shipped inside the probe.
Snapshot the sampled rows before calling the block. Advanced indexing
copies, so the snapshot is not a view onto src.
Regression test pins it: a block mutating img in place must still report
|delta| = 2x|before|. Verified it fails with 0.0 against the old code.
Also: the readout used toFixed(4), which hides exactly the range these
values live in — the intervention map measured block activity varying in
the 4th decimal, so 0.0000 was ambiguous between "tiny" and "nothing".
Now toPrecision(3).
Detail probe: records the seed with each run, so a low `structure` can
say which of its two causes applies — different seeds (two separate
generations, never an A/B; a reference pins the subject, not the sample)
or the same seed whose shot the change actually moved, which is a
result. The numbers cannot separate those; the seed can.
h3_block_repeat runs the named block(s) on their own output before
passing it on — a second pass that never leaves latent space. To first
order this doubles that block's contribution; the nonlinear part is a
genuine second refinement of its own result. ~2% of a step per repeat.
detail_probe scores whether that (or anything else) actually sharpened
the picture, by comparing a run against the previous one on three
numbers:
detail high-frequency energy — did fine structure increase
structure low-frequency survival — same shot, or a different one
edge-aligned did the EXTRA detail land on existing edges, or spread
The third is the one that earns its place: noise is high-frequency too,
so `detail` alone cannot tell sharpening from grain — and grain is
exactly what an out-of-distribution stream would produce. Verified on
video-latent-shaped data: unsharp-mask reads detail x3.04 / edge +0.885
("sharper"), uniform noise reads detail x1.12 / edge -0.248 ("grain").
Both raised detail; only edge-alignment separated them.
Measured in latent space deliberately — the tensor is already in hand at
the call site DynaShift banks from, so no VAE decode and no new
plumbing. A real approximation, valid for the A/B it is for (same seed,
identical decoder both sides) and documented as ratio-only.
Reuses block_influence's collection switch rather than adding a second
research toggle, and pairs CONSECUTIVE runs, which is the workflow
anyway: generate with it off, generate with it on, read the row.
Also scores second_pass sharpen, quality_sharpness and segmented
detailing — every mechanism here that claims detail and is currently
judged by eye.
previous one didn't
Magnitude alone cannot distinguish a block contributing something nothing
before it contributed from a block pushing harder in the direction the
last one already went. Records cos(f_i, f_i-1) on the same strided video
rows: ~1 means it amplifies its predecessor (little new per block), ~0
means orthogonal (a genuinely new component), negative means it partly
UNDOES the previous block — which real stacks do, and which would mean
the effective work is far smaller than the sum of the magnitudes.
This is the number that speaks to whether 50 blocks are doing 50
different things or one thing 50 times.
Step boundaries are respected: a block index that does not exceed the
last one seen starts a new step, so block 0 is never compared against
block 49 of the previous step — that pairs the end of one denoise step
with the start of the next, which is not a depth relationship.
Reads out in Refinement & Taste beside flatness, with a plain-language
verdict and a per-block column. Verified live on seeded data: amplifying
(+0.81), undoing (-0.49) and orthogonal (+0.06) regimes all round-trip
through the API and render. Seed data cleared.
Research data belongs in Settings > Refinement & Taste — readable,
downloadable, clearable, with its own collection switch — not in a run
report that already carries more than anyone but us needs.
- Removed the sampler widget and its Editor knob entirely. Collection is
driven by a global switch (env var live for this process + on-disk copy
so a fresh rental comes back recording), the same two-layer mechanism
trajectory_probe uses.
- Removed the per-rating console readout. Findings go to the panel;
failures still speak through funpack_log.
- New section in Refinement & Taste: toggle, recorded count, Download,
Clear, a plain-language flatness verdict, and the busiest blocks ranked
with their liked-vs-disliked delta.
- Routes: GET/POST /api/block_influence, /export, /clear.
The switch gates whether the PROBE is installed and nothing else —
save_pending/commit stay unconditional, with a test pinning that, since a
toggle that quietly stops recording ratings is a mistake this project has
already made once.
Verified live: routes round-trip, the panel renders a seeded 6-run
profile, and a planted rating-correlated bump at block 31 surfaces as
+0.0107 in the UI. Seed data cleared afterwards.
Records ||f_i(x_i)|| / ||x_i|| on video rows for every block, averaged
over each run's steps, and pairs the profile with that run's rating.
Answers the question that gates rating-driven block weighting: is H3's
depth profile flat (nothing to aim at) or structured? Reports a
scale-invariant flatness number, and once a key has 2+ liked and 2+
disliked runs, which blocks run hotter on the liked ones.
Steers nothing. Installed BEFORE any steering hook so the delta it
records is the block's own, not one already containing an injection.
Bounded cost: two norms per block per step on ~512 strided rows, no
extra forward pass, results accumulated as device tensors and synced
once per run rather than 50x per step. Says so explicitly when it
records nothing, since an empty capture and a flat profile are
otherwise indistinguishable downstream.
Adjacent to per-block residual steering, removed in b7ed4f9 for
finding no rating signal above the 4th decimal -- that pass had no
per-step resolution, which its own removal note named as the gap.
This either reproduces that null definitively or it doesn't.
the negative one
DynaShift was forced off for H3 in the 2026-09-03 disable pass on
theoretical grounds (LTXAV-calibrated, never actually tested on H3) --
re-enabled here to find out. Also gives it a second bank: negative_memory
now stores liked generations the same way it already stores disliked
ones (is_positive_profile, same sign-only convention h3_repr_steering
uses), and _build_dynashift_wrapper computes mean(liked) - mean(disliked)
over each bank's matching-resolution frames as a pull direction --
deliberately NOT nearest-neighbor matching toward one liked example,
which would risk reproducing that specific past generation. Needs
MIN_BANK_FOR_PULL (2) entries in each bank at the run's own latent
resolution before it contributes anything.
Reuses the existing dynashift/dynashift_strength/dynashift_threshold
controls for both directions -- no new Editor UI needed.
Never worked on this R2V-wired setup (prompt_hash always came back
None, silently disabling the guard on every sweep) and fixing it was
explicitly declined. Removed rather than left inert -- a guard that
never engages is worse than no guard, since it implies a protection
that was never actually there.
Misread "disable that .pt tracking" as "stop recording ratings" --
the actual ask was about console noise, already handled by removing
the block-sweep print. This toggle instead silently stopped every
rating from being recorded while it was on, confirmed live: a real
key sat at "0 liked / 0 disliked" through multiple ratings with no
visible error, and the "committing weight" log line looked correct
the whole time (fixed separately, but the toggle itself should never
have existed).
Removed cleanly rather than patched: capture_enabled/set_capture_enabled,
the enabled gate in save_pending/commit, the /enabled route, the API
client method, and the Settings toggle row. save_pending and commit
now behave exactly as they did before this was ever added -- every
rating is recorded, unconditionally.
"Committing weight +X" printed unconditionally, before checking
whether there was a pending capture to attach it to -- so a correctly
computed reward with nothing to pair it with (capture was off, or the
sampler's h3_repr_steering toggle was off, on the run being rated)
looked identical in the log to a real commit. That's the gap that
made "0 liked / 0 disliked" on a cleared key impossible to diagnose
from the console alone: the weight line looked right every time.
commit() now returns "recorded" | "no_pending" | "disabled" | "no_key"
and the caller prints which one actually happened.
Picker: h3_repr_steering only ever reads the reward's SIGN now (see
the binary-direction commit), so a 1-10 scale was granularity nothing
downstream uses. Down to two buttons -- reuses the scale's own "10"
and "1" endpoints, zero backend change.
REINS capture toggle: independent of the sampler's h3_repr_steering
widget, which controls whether the learned direction gets APPLIED --
this new per-key flag controls whether it keeps RECORDING at all, so
data collection can be paused without touching the steering setting
or discarding what's already captured. Settings > Refinement & Taste
gets the toggle; save_pending() is the single gate.
Weighting by rating magnitude (a 9/10 pulling harder than a 6/10) is
the better estimator in the limit, but needs enough data for a
weakly-rated row's noise to average out. At the handful of ratings a
key actually accumulates, a mediocre "6" just gets a vote and dilutes
the direction with an ambiguous case instead of being excluded by it.
Binary spends what little data exists on the clearest examples only.
block_sweep(), multi-block capture, and prompt-hash grouping are
unaffected -- only direction()'s own math changed.
Captures every candidate block (5,10,...,45, plus DEFAULT_BLOCK=25)
read-only alongside the one that actually steers, so block_sweep()
can answer "which block really separates liked from disliked" from
data already being collected -- reusing trajectory_probe's own
permutation test (proven correct on this project's real data) instead
of trusting REINS' 50%-depth heuristic (their own paper picks the
best layer per model by sweep, not by applying one universal
percentage) or porting a diagnostic built for a different question
(LTXAV's identity-block search measured cross-scene consistency, not
liked-vs-disliked separation).
Grouped by prompt_hash (trajectory_probe's own, reused) so cross-
prompt content differences can't masquerade as a rating signal -- the
same confound that inverted the first cross-prompt read of this
project's output value function.
Settings > Learning > Refinement & Taste gets a per-key REINS panel
(status, block-sweep table, export/clear) mirroring the trajectory
probe UI, plus matching /api/h3_repr_steering routes -- so data can
be pulled off a rental and handed over directly instead of relayed
through pasted console logs.
Root cause of "every rating counts as liked": the 'rating' node
input is a COMBO widget the graph builder validates against
V2_RATING_LABELS (builder.py's _live_value) -- the same guard that
falls back to a node's default when a saved LoRA/model choice isn't
installed on this machine. The H3 scale sends bare digit strings,
which were never added to that list, so EVERY H3 rating since the
scale shipped was silently rejected as "not a valid choice" and
replaced with the node's declared default, "Missing action" (+0.05)
-- explaining the diagnostic log exactly: legacy_score=None (never
reached the digit-parsing path at all) and reward +0.050 (the
literal default, not a mapped value).
Confirmed with the log line the user captured: the fix earlier this
session to normalize_refiner_v2_rating's digit math was correct and
irrelevant -- the real rating value never reached that function.
Fix: add "1".."10" to V2_RATING_LABELS so they pass validation
unchanged, same list both rating widget declarations already read.
H3's only validated (or being-validated) rating consumers are
h3_phrase_emphasis and h3_repr_steering. Everything else --
value_guidance/absolute taste steering, output_guidance,
trajectory_guidance, dynashift, embed_guidance, score_slider,
taste_nearest_prompt, and the path planner -- was built and
calibrated for LTXAV or for a value function nothing on H3 reads.
None of it helps REINS; it only trains value functions that are
never consumed and prints status lines for mechanisms that can
never fire, which is what actually produced the "80% of the log is
useless" bug report this responds to.
Forced off at two chokepoints (top of sample() in samplers.py once
self._is_h3 is known; top of refine_v2 in conditioning.py once the
clip's family is known), not just hidden in Engine Settings, so a
project saved before this still behaves correctly. Value-function
and Absolute-taste TRAINING (not just application) is skipped too --
per-scene, per-key, and the main path -- since nothing on H3 will
ever read what they'd produce.
normalize_refiner_v2_rating's legacy digit path snapped a bare number
to the nearest existing category label and used THAT label's reward
verbatim. Missing action's +0.05 was calibrated for its own meaning
("the frame was fine, just no motion") -- not "a mediocre score" --
but 5 and 6 both landed in that bucket, so rating something a 5/10
produced a POSITIVE reward. Every h3_repr_steering weight built from
those ratings was backwards.
Reward is now a straight linear map across the full 1-10 range
(1 = -0.9, 10 = 1.0), so the sign matches what the scale promises.
Level/missing_axes/skip flags still come from the nearest label --
harmless, since nothing needs them exact for a legacy score.
Every rated run now pulls proportionally to its own reward instead of
being bucketed into two equal-weight groups -- a 9/10 counts for more
than a 6/10, instead of both counting as one full vote. Centring the
WEIGHTS (not the descriptors) before the weighted sum is what makes
shared content cancel exactly even when the weights are unbalanced;
see the module docstring for the identity. This is what actually
justifies the scale existing at all, rather than just picking one of
two buckets with extra steps.
Also: the H3 scale's heart/loved button did nothing --
normalize_refiner_v2_rating only applies the loved boost inside the
branch that matches an existing category label, and a bare number
never reaches it (confirmed: '7' and '7|loved' produced the identical
reward). Removed it from the scale rather than wire a boost with no
established meaning on a 1-10 input.
The category grid assumes axis-specific effects (phrase emphasis,
dynashift...) this project has not proven land differently on H3.
Replaces it with a single scale for H3 pipelines only -- the backend
already turns a bare number into the nearest existing rating profile
(normalize_refiner_v2_rating's legacy-score path), so every consumer
(value functions, phrase emphasis, dynashift, path planner) keeps
working unchanged; only what the picker sends changes.
Also fixes h3_repr_steering: it was splitting liked/disliked by the
sign of the blended scalar reward, which counts near-misses like
Missing action (+0.05) as "liked" even though that reward exists for
the value functions' quality-landscape ascent, not as a liked/disliked
verdict. Now filters by rating KEY (LIKED_KEYS/DISLIKED_KEYS) and
excludes everything ambiguous instead of guessing a side for it.
Reaches inside the forward pass instead of around it: captures the
video rows of one block's hidden state per generation, and once a
refinement key has 3+ liked and 3+ disliked runs, adds their mean
difference back into that block's output on every later run. No
gradient, no MLP -- a mean-difference of vectors is the whole model.
Masked to video rows only via the block's own mod_segments tags,
never audio or text -- the one documented failure of this class of
intervention (project_self_consistency_failed.md) happened on a
different model (LTXAV) via joint audio-video attention.
Off by default, H3-only, unvalidated.
0 = only cemented phrases get extra attention. 1 = that bias
collapses to nothing, letting untrained content compete evenly.
Same attention-bias channel as the existing toggle, scaled down.
Settings ▸ Learning ▸ "Start a fresh measurement". Deleting the refinement key
was not enough: the probe log lives outside the <key>.* sidecar convention on
purpose, so it survives a key clear, and the per-bucket heads are REBUILT from
it on every rating — the next rating restored them from runs still on disk.
So the clear takes both, and leaves the switch alone: "throw away what was
measured" and "stop measuring" are different requests.
Verified against a running ComfyUI: 12 runs and the heads file removed, count
back to 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ComfyUI executes a node under torch.inference_mode(), so a descriptor read out of
the probe log becomes an inference tensor and autograd refuses it:
Trajectory probe intake failed: Inference tensors do not track version counter.
Every head stayed empty and the run reported "nothing learned yet" on every
generation — so the feature was OFF while looking merely untrained, which is the
one failure invisible from the outside: "not steering" is also what correct
behaviour looks like before enough ratings. Seen on a live rental with 73 runs
logged and zero of them trained.
The output value function's trainer already carries this guard; this one did not.
Also: the bucket-isolation test was measuring TRAINING accuracy, where a
few-hundred-parameter MLP fits 20 noise vectors perfectly and every bucket scores
1.0. It now holds rows back, which is the only version that can tell a head that
learned from one that memorised.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every rating-driven mechanism shares one gate that opens only over the last half
of the schedule — after motion and layout are decided — so a rating about
movement has never had a window to act in.
This learns one value function per schedule quarter, each trained on that
quarter's own predicted output, each steering inside its own window, with no
shared late ramp. It reads the probe's log, so it starts trained on whatever has
already been rated.
Basis: 51 rated runs across two independent sessions, different prompts and
balances. The early window measured 88% and 89% of the late window's effect,
direction unanimous across all 8 bucket-by-session results, sessions combining
to p 0.043. Cross-validated on the clean session the heads rank held-out runs at
AUC 0.73-0.82.
Beside output_guidance, not replacing it, and off by default: the honest test is
the same seed with it off and on. A window with under 10 samples is byte
identical to the feature being off, so "nothing changed" stays readable, and the
run says which windows can steer rather than only that it is on.
Not included: axis routing (quality late, action early). It is the next step and
it would muddy this A/B.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Runs of one prompt share most of their content: on a real 21-run session, 86% of
each descriptor was the part every run had. Comparing raw vectors measured that
shared part, so every distance came out near zero and near equal — mean cosine
0.97, and no separation at ANY bucket including the late ones where steering
demonstrably works.
Centring first flips the same data from -0.002 (backwards) to +0.03 and p 0.83
to 0.11, consistent across all four buckets. Not significant at 5 good / 16 bad,
but pointing the right way and now limited by sample size rather than by the
statistic.
The descriptor was not at fault. adaptive_avg_pool1d over the flat latent keeps
512 REGIONAL means, not one global mean, and it distinguishes clips fine.
Also: the prompt id is quantised before hashing. Raw bytes answered "are these
bit-identical", so any value crossing a rounding boundary changed the id — one
prompt produced 20 distinct ids across 21 runs. Stable now for jitter of the
size a GPU produces; no exact hash of floats can do better.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Settings ▸ Learning gains Download and Load. A rental is replaced whenever
something breaks, refinements/ is not in git, and the reading needs ~16 rated
runs — so without a way off the box the count restarts on every instance and
the answer is never reached. 16 runs is ~85 KB.
Importing the same export twice adds nothing: the permutation test reads n
directly, so duplicated runs would report a separation as far more significant
than the evidence supports.
An uploaded .pt is a pickle, so the restore parses it with weights_only — every
field in a run is a tensor or a primitive, and anything else is refused rather
than executed. Nothing but numbers, a rating key, a timestamp and a one-way
prompt hash is stored; no prompt text has ever been written to these files.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Settings ▸ Learning ▸ Trajectory probe: a switch to record, the run count,
and a Read button that gives the answer in plain language. The switch is the
same environment variable the sampler reads and the editor sets it in the
running process, so it takes effect on the next generation with no restart —
and it is written to disk, so a restarted box comes back recording instead of
having quietly stopped.
tools/trajectory_separation.py stays as the terminal door for a headless box.
Verified against a running ComfyUI: toggle both ways, persistence across a
real restart, and the reading recovering a planted early signal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every rating-driven mechanism shares one gate (_make_steer_ramp) that opens
only over the last half of the schedule, so a "Missing action" rating has no
window in which it could act. Banking x0_hat per schedule position would give
it one — if the early half carries any signal. This measures that.
Off unless FUNPACK_TRAJECTORY_PROBE=1. Records, never steers. Built for H3:
buckets clamp to few-step runs, batched predictions are refused rather than
averaged with their unconditional row, and each pass binds its own schedule.
Read the result with tools/trajectory_separation.py.
fix: reference numbering counts within its own kind
A reference slot resolves per kind, so marking an image then a video labelled
the video R2 while the only slot reaching it was "Reference video 1" — there
was no way to have an R1 video and an R1 image at once.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Prompt repair lost most of itself in an earlier stage; what was left was a
fallback that appended missing phrases back onto the prompt, reachable only
through a list nothing fills any more. Gone. The axes stay — active_repair_axes
is the "what's missing" signal the advisor reads, and it never edits a prompt
by itself.
prompt_repair and prefer_wired_conditioning are read by nothing at all and only
survive in old saved settings, so the card stops printing them.
Bounded attention is built and has never been run, and its left/right split is
fixed. Both labels now say so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
digital-garbage