STAS (arXiv 2603.17825) sets massive-activation channels on frame 0 and
frame edges to alpha x peak at one block over the first 40% of steps.
Rating-learned dials now share rated_dial.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Skip guidance at CFG 1: the weak copy reuses blocks before the branch and
skips the branch block, so only the tail runs twice. Picture only; capture
hooks ignore the weak pass; compiler off while installed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Shot memory reuses a liked gen's coarse starting noise (layout) with fresh
fine noise; reuse-vs-fresh, which shot and how strongly are learned.
Decisiveness is Temporal Score Rescaling (arXiv 2510.01184) on the picture
rows, k learned per key or set by hand.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Studio gains an enhanced_prompt output; the link wires to it. Enhancer off or failed sends prompt + postfix unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ComfyUI tokenizers drop an unknown system_prompt kwarg, and Gemma skips its chat template by default. Instructions now go in the user turn of the model's own template, which also places the image.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- region lock: layout patch matched the old PackedLayout signature and crashed every H3 run; interior pins now use upstream's native any-frame placement
- REINS/Q-steer/AV decouple/probes: masks cover the target video/audio rows only, not pins, refs or vision pads
- video detail: works with the PDD FinalLayer signature; now in Engine Settings
- Q-steer: steers queries before RoPE, not after (old banks are not reused)
- ALG anchor: blurs the frame-0 pin on H3
- temporal style: hidden on H3
- shadow negative: main-window chip names what it switches off without compose
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Own tab with its full sampling settings (temperature, top-p, top-k, min-p,
repetition/presence penalty, seed, greedy) and a last-run before/after view.
Studio's source_image now reaches the enhancer on the per-scene path too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
These were force-disabled for a gate bug, not a measured failure: they
all rode _make_steer_ramp's last-half window, which measured 0-of-4 and
0-of-12 real step coverage on H3's shift-based schedules because it read
sigma as schedule position, which is false there. That gate was already
rewritten to recover the true position from the schedule's own base grid
and is wired into all three call sites -- their correction steps are
also self-normalizing (a fraction of the live tensor/embedding's own
norm, not an LTXAV-tuned constant), and trajectory_guidance's buckets
were built for H3's schedules from the start. Re-enabled to actually
test them now that the thing that broke them is fixed.
embed_guidance/taste_nearest_prompt stay off: different value function,
and phrase emphasis's own manipulation check directly measured
near-invisible pushes from it at H3's calibration -- an actual negative
result, not a stale gate.
Unvalidated on real hardware -- this removes a stale reason for
disabling them, not a claim that any of the three help.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
explore_first_step reads the same output-space value function
output_guidance trains, but training/snapshotting was gated off for H3
entirely -- a decision made before explore_first_step existed, when
output_guidance (still correctly forced off there) was its only consumer.
Result: the value function could never reach MIN_SAMPLES on H3 no matter how
many generations were rated, so explore_first_step printed "not ready yet"
forever. Un-gate the snapshot save and the training call for H3; leave
output_guidance's own application (and the separate conditioning-payload VF
for embed_guidance/value_guidance) forced off as before -- that decision is
unaffected.
No new automated test: no existing fixture exercises this rating-commit code
path (FunPackVideoRefiner.run's post-scene training block) cheaply enough to
extend honestly; verified by full suite pass (1491/1491) and direct tracing
that no other call site re-introduces the gate.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Corrected the actual design: encode two FULL conditionings ($style
SHORTCUT1 $postfix and $style SHORTCUT2 $postfix) and average them row by
row -- (1-strength)*A + strength*B per position, strength=0.5 giving
exactly (A+B)/2 as originally specified. The prior mean-vector version
collapsed phraseB's whole sentence into one flat direction applied
identically everywhere, discarding its own word-by-word content.
Resamples the shorter side (linear interpolation along the token axis)
when the two shortcuts tokenize to different lengths, so both sides line
up for the average without changing the output's token count.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Confirmed on the rental: the previous version (phrase-scoped shift) showed
zero visible effect even at strength 1.5. Qwen's rows are contextual, so
every word in "a cat sits on a mat" already carries "sits", not just the
verb's own row -- editing 1-4 rows out of the sentence left the rest still
pointing at phraseA's action. Now shifts the whole prompt-token range toward
the alt sentence's own whole-prompt mean.
Multiple [a|b] spans in one prompt are additive: each span's delta is
measured against the original conditioning, not against a prior span's
already-shifted result, so they don't compound order-dependently.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Default 0.5 shift left phraseA dominating the video. [phraseA|phraseB:N]
overrides strength per blend. Fixed a shared regex gap along the way: every
weight/strength markup in this file (not just blend) silently rejected
leading-dot decimals like ".8" -- fell through to literal text with no
error, reproducing the exact "override did nothing" bug this feature exists
to fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adversarial review of the previous commit found two real leaks and fixed
them: a colon in an alt phrase reached the real text encoder unstripped
(no bracket survives to re-catch it after substitution), and a wired
conditioning re-blended a second time because the re-tag guard only
checked for tag absence. Also reports per-span placement failures instead
of only counting successes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A latent unsharp on H3's 24-channel latent does not translate to picture
sharpness. Wrong mechanism, not wrong number; removed rather than tuned.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The parser existed but nothing called it; the syntax went to Qwen as text.
One pass strips both markups so timed and weighted spans share one clean text;
both encode paths carry them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The editor strips the brackets before the encoding node sees them and hands the
windows to Studio as char spans on the text it gave that node.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Query rows are attended in chunks split at window edges, each with its own key
bias: exact, no SxS mask, no extra model call. Outside the window the phrase is
masked from video rows (H3-World single-egress); audio and text rows never see a
window. Composer lists the windows under the prompt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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).
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.
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.
Duplicated what Settings > Refinement & Taste already computes and
shows as a table on demand. Console still gets the compact one-line
commit confirmation; the full per-block permutation-test ranking now
lives only in the panel.
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.
Both were captured "regardless of the toggle" by original design (so
turning guidance on later works immediately), which meant they kept
running on H3 even after output_guidance/dynashift were forced off —
exactly the kind of function the "disable everything unrelated to
REINS" pass was supposed to catch, missed because the capture point
predates any toggle check.
Confirmed live: the user's own log showed "DynaShift negative bank
updated — 1 entry" after the rest of the disable pass had already
landed.
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.
