"Training is slow" and "training is slow ON THE CPU with four million
parameters" are different problems with different fixes, and nothing in the log
told them apart — leaving it to be inferred from timings, which is how the
first guess at this went to the disk instead.
One line per run per value function, naming the device and the parameter count.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Off by default. On, the timeline sits as a strip — its header, so the title,
transport and Composer stay reachable — and opens to full height while the
pointer is on it or something inside it has focus.
It also opens for the whole of a DRAG. During one the pointer is in drag mode
and a strip that stays shut is a drop target you cannot hit, so dragenter holds
it open and drop/dragend closes it. The tracker counts depth rather than
holding a boolean: dragleave for the element being left fires AFTER dragenter
for the one being entered, so a flag closes the strip mid-drag as the pointer
crosses from one clip to the next.
The strip height is MEASURED from the header rather than fixed: .zone-head
wraps when the zone is narrow (34px min-height, 138px on a normal window), so
any constant clips the very controls the strip exists to keep visible. It is
re-measured on resize. flex: 0 0 auto matters — .zone is a shrinkable column
child, and a height with min-height: 0 let the workspace squeeze the strip to
1px. The body collapses rather than being display:none so the lanes are not
torn down and rebuilt on every open.
A browser preference like the colour scheme, not a project one, and registered
in the same Settings section for that reason.
Also in this commit:
- vf: train wherever there is a GPU, not only when the sample sits on one. The
x0 value function's snapshot loads with map_location="cpu", so keying on the
sample left it training on the CPU — the run that named it measured
"x0 vf: train 10.1s" against 46ms for the same net locally.
- The stage line counted nested sub-stages twice and reported 153% of a run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Measured rather than assumed. The three value functions are ~16s of a 26.5s
rated run, and the phase that costs it is the training, not the disk:
text cond, dim 5120 3,982,467 params 16.2MB load 19ms TRAIN 1068ms save 26ms
x0 latent, dim 512 221,955 params 0.9MB load 3ms TRAIN 46ms save 2ms
The nets are tiny in FLOPs and not in parameters — a 5120-wide input makes the
first layer 94% of the weights — and twenty Adam steps over that ran on the
CPU because the checkpoint loads with map_location="cpu". Roughly 1.1s here,
about 7.5s on the rental's vCPU, three times per rated generation.
The conditioning is already on the GPU when there is one, so training moves
there and moves back before returning: the buffer and the checkpoint stay on
the CPU so the file remains portable, and inference elsewhere still finds a CPU
module. Adam's state does not follow module.to(), so it is dropped on the move
rather than migrated — it is rebuilt lazily and never persisted, so a run
starts from fresh moments either way.
Nothing about what training does changes; only where it runs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both value functions cost about eight seconds a run, and one has ten times
fewer parameters than the other — so it is not the training math, and the whole
per-call cost is somewhere else. The key file writes at roughly 3 MB/s on the
rental, which makes the checkpoint write the first suspect, but that is a guess
until the three phases are timed apart.
Also times loading the x0 snapshot, which is a full packed latent read from the
same disk.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replace absolute reward regression with Bradley-Terry pairwise ranking
(DPO-style) so the value function learns relative preference between
sampled pairs rather than fitting noisy absolute reward floats.
Add a 3-member bootstrap ensemble; gradient() and ascend() scale by
inter-member agreement so guidance backs off in regions the ensemble
hasn't seen data for, instead of confidently extrapolating (the
static-character collapse this was meant to guard against).
Loosen the "Wrong but loved" exclusion in conditioning.py so a strong
loved_modifier signal is no longer filtered out by skip_value_function
- "Wrong but loved" and "Missing but loved" now train the value
function, per the stronger-signal precedence requested.
Legacy single-net checkpoints migrate transparently on load (weights
remap into all ensemble members, rating buffer preserved intact).
New LatentValueFunction (value_function.py) mirrors OnlineValueFunction but is
trained on a pooled snapshot of the sampler's own predicted output (x0_hat)
instead of the input conditioning, borrowing NoiseTilt's core idea (reward
gradient from the model's own prediction) without its SDE noise-term trick,
which has no analogue in this deterministic-ODE sampler.
New output_guidance / output_guidance_strength toggle on the Chain Sampler
(off by default): per-step correction installed as the outermost
model_function_wrapper (on top of embed_guidance/score_slider), same
near-zero-forward-pass shape as embed_guidance. End of run writes a small
pooled snapshot per refinement key; the next rating cycle trains the value
function on it via a new call site in conditioning.py.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Route the conditioning variant from per-config outcome history, via the one
charter-safe surface: the INTERNAL value-function ascend+search (not user knobs,
so no WYSIWYG / hidden-settings violation). Sampler-owned visual ops stay on the
sampler (boundary law kept).
Directly serves the original complaint (Perfect not locking, Awful reproducing):
- _v2_path_conditioning_plan: matching-arm Beta posteriors -> liked-rate mean ->
"lock" (>=0.6: hold the liked conditioning, skip the stochastic search),
"explore" (<=0.4: widen the search to flee a disliked region), else "normal"
- value_function.search gains eps_scale to widen/tighten the search radius
- threaded conditioning_plan through ascend -> apply_scene_refinement_keys ->
finalize; computed once per run from path_outcomes + prompt concepts; logged
Only active once the per-key VF is ready, since it rides the existing ascend path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
search() generates 16 random unit-direction perturbations (2.5% of
conditioning norm each), scores each with the VF, returns the best.
Runs after ascend() in _v2_ascend_conditioning — ascent exploits the
known gradient direction, search explores the neighbourhood around it
for any better nearby region. Total cost: 16 VF forward passes ≈ free.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
LTX packs all spatial/temporal dims into a single token [1, 1, C*T*H*W]
during sampling. Per-step latent steering requires hooking into the model
internals, not the k-diffusion callback. Removed: compress_latent(),
_load_latent_value_function(), latent VF training, motion floor callback,
latent_value_guidance toggle in Studio.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Unconstrained ascent was stripping action content from conditioning to reach
a "safe" static-character attractor. Added 8% displacement cap relative to
original conditioning norm — quality direction is still steered but the
prompt content can't be erased. Also raised MIN_SAMPLES from 5 to 10 to
reduce overfitting risk before the value function activates.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Trains a second OnlineValueFunction on rated video latents (compressed via
channel mean-pool). At each denoising step, computes ∂reward/∂latent through
the lightweight MLP and applies it as an in-place drift to the noisy latent —
gradients never touch the 22B transformer. Controlled by embed_guidance_strength
alongside the existing conditioning guidance. Studio toggle: "Latent guidance".
Cleared on session reset alongside the conditioning VF.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Before each denoising run, when the value function is ready, gradient-ascend
the text conditioning until predicted reward plateaus (self-terminating: step
size = 0.5% of conditioning norm, stops when delta < 1e-3, caps at 50 steps).
No user-facing params — fully automated. Per-step wrapper still runs on top.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
.clone() on an inference tensor inherits the inference flag even inside
inference_mode(False). torch.empty() inside that context creates a true
normal tensor; copy_() fills it from the inference tensor without flag
inheritance.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
MLP lives on CPU, conditioning on GPU — device mismatch silently fell back
to fixed direction. Now moves conditioning to MLP device for the forward/
backward, then returns gradient to original device.
Also surface value function fallback errors instead of swallowing them.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Buffer tensors were created inside inference_mode — detach().cpu() doesn't
escape the inference flag. Wrapping the whole method ensures the clone and
all subsequent tensor ops produce normal autograd-tracked tensors.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
torch.enable_grad() cannot override torch.inference_mode() in older PyTorch.
Add inference_mode(False) as outer wrapper in both train_on and gradient.
Also .clone() tensors from buffer/input before grad tracking to avoid
inference_mode tensor view restrictions.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ComfyUI execution context may have torch.no_grad() active, which kills
gradient tracking needed for training and gradient computation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
value_function.py: OnlineValueFunction MLP that trains incrementally on
(conditioning, reward) pairs. Uses replay buffer of 100 samples, 20
gradient steps per update. save/load to .pt file via refinement_state_path.
Studio popup: new "Value guidance" toggle. When on, trains the value function
after every rated generation (good and bad). Falls back to no-op if conditioning
deserialization fails.
chain sampler: _load_value_function reads the trained MLP if ready (5+
samples). _build_embed_guidance_wrapper uses value function gradient
(∂reward/∂conditioning) when available, fixed liked direction as fallback.
Log reports which mode is active and sample count.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
