digital-garbageandClaude Opus 4.8
8725b9d517
The score_slider synthesized a symmetric +/-liked pair, so its axis was "liked vs anti-liked mirror" — meaningless as an away-from-bad signal. When 3+ disliked/awful gens have populated the existing bad_dir slot in the same taste store, build the slider's minus pole from that real disliked direction instead. The (eps_+ - eps_-) axis becomes good-vs-bad, actively steering the noise prediction away from the conditioning region that produced rated-bad gens. Falls back to the symmetric mirror when no bad_dir exists. Reuses existing plumbing: bad_dir is already accumulated identically to liked_dir (running-averaged unit-norm delta from session mean). Generalized the direction loader to _load_taste_direction(slot); load + pass bad_dir at the call site; log/run-mechanism note which pole is active. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
115 lines
4.7 KiB
Python
115 lines
4.7 KiB
Python
"""Unit tests for the FreeSliders-style score-space taste slider on the
|
|
Scene Chain sampler (_build_score_slider_wrapper).
|
|
|
|
Covers the 3-pass combine eps_mod = eps_base + eta*(eps_+ - eps_-), the high-sigma
|
|
base-only warmup, the eta=0 no-op, and that it leaves a clean wrapper to restore.
|
|
"""
|
|
import sys
|
|
import types
|
|
from pathlib import Path
|
|
|
|
import torch
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
|
|
|
# Minimal comfy stubs so `import samplers` works without a full ComfyUI env.
|
|
for _name in (
|
|
"comfy", "comfy.k_diffusion", "comfy.k_diffusion.sampling",
|
|
"comfy.model_sampling", "comfy.nested_tensor", "comfy.sample",
|
|
"comfy.samplers", "comfy.utils",
|
|
):
|
|
sys.modules.setdefault(_name, types.ModuleType(_name))
|
|
sys.modules["comfy.nested_tensor"].NestedTensor = object
|
|
|
|
import samplers # noqa: E402
|
|
|
|
N = 10 # packed latent length used by the fake model
|
|
|
|
|
|
def _make_wrapper(eta, monkeypatch, bad=None):
|
|
s = samplers.FunPackLTXAVSceneChainSampler()
|
|
# Single-stream: audio protect is an identity passthrough.
|
|
monkeypatch.setattr(s, "_protect_audio", lambda steered, original: steered)
|
|
model = types.SimpleNamespace(model_options={})
|
|
liked = torch.ones(16)
|
|
old = s._build_score_slider_wrapper(model, liked, eta=eta, bad_dir=bad)
|
|
return model.model_options["model_function_wrapper"], old
|
|
|
|
|
|
def _apply_fn(inp, ts, **c):
|
|
"""ε filled with the mean of the conditioning -> linear in c so we can predict
|
|
the symmetric combine."""
|
|
cc = c["c_crossattn"]
|
|
return torch.full((1, 1, N), float(cc.float().mean()))
|
|
|
|
|
|
def _args(sigma, cond):
|
|
return {"input": torch.zeros(1, 1, N), "timestep": torch.tensor([sigma]),
|
|
"c": {"c_crossattn": cond}}
|
|
|
|
|
|
def test_score_slider_returns_no_previous_wrapper_when_none(monkeypatch):
|
|
_, old = _make_wrapper(2.0, monkeypatch)
|
|
assert old is None
|
|
|
|
|
|
def test_score_slider_warmup_is_base_only_at_high_sigma(monkeypatch):
|
|
wrap, _ = _make_wrapper(2.0, monkeypatch)
|
|
cond = torch.randn(1, 4, 16)
|
|
out = wrap(_apply_fn, _args(0.9, cond)) # ramp = 1 - 0.9*2 < 0 -> base only
|
|
assert torch.allclose(out, torch.full((1, 1, N), float(cond.float().mean())))
|
|
|
|
|
|
def test_score_slider_applies_in_quality_phase(monkeypatch):
|
|
wrap, _ = _make_wrapper(2.0, monkeypatch)
|
|
cond = torch.randn(1, 4, 16)
|
|
base = torch.full((1, 1, N), float(cond.float().mean()))
|
|
out = wrap(_apply_fn, _args(0.1, cond)) # ramp = 0.8 -> 3-pass combine
|
|
assert not torch.allclose(out, base)
|
|
# liked dir is all-positive and step is symmetric, so eps_+ > eps_- -> push is positive
|
|
assert float(out.mean()) > float(base.mean())
|
|
|
|
|
|
def test_score_slider_eta_zero_is_noop(monkeypatch):
|
|
wrap, _ = _make_wrapper(0.0, monkeypatch)
|
|
cond = torch.randn(1, 4, 16)
|
|
base = torch.full((1, 1, N), float(cond.float().mean()))
|
|
out = wrap(_apply_fn, _args(0.1, cond))
|
|
assert torch.allclose(out, base)
|
|
|
|
|
|
def test_score_slider_strength_scales_push(monkeypatch):
|
|
cond = torch.randn(1, 4, 16)
|
|
base = float(torch.full((1, 1, N), float(cond.float().mean())).mean())
|
|
wrap1, _ = _make_wrapper(1.0, monkeypatch)
|
|
wrap3, _ = _make_wrapper(3.0, monkeypatch)
|
|
push1 = float(wrap1(_apply_fn, _args(0.1, cond)).mean()) - base
|
|
push3 = float(wrap3(_apply_fn, _args(0.1, cond)).mean()) - base
|
|
assert push3 > push1 > 0
|
|
assert abs(push3 - 3.0 * push1) < 1e-4 # linear in eta
|
|
|
|
|
|
def test_contrastive_bad_pole_replaces_symmetric_mirror(monkeypatch):
|
|
# When bad_dir == liked_dir, both poles step the SAME way, so eps_+ - eps_- collapses
|
|
# to ~0: the contrastive minus pole genuinely replaces the -liked mirror (which would
|
|
# still push). Proves the bad direction, not a mirror, defines the minus pole.
|
|
cond = torch.randn(1, 4, 16)
|
|
base = float(torch.full((1, 1, N), float(cond.float().mean())).mean())
|
|
sym, _ = _make_wrapper(2.0, monkeypatch) # symmetric +/-liked
|
|
contra, _ = _make_wrapper(2.0, monkeypatch, bad=torch.ones(16)) # bad == liked
|
|
sym_push = float(sym(_apply_fn, _args(0.1, cond)).mean()) - base
|
|
contra_push = float(contra(_apply_fn, _args(0.1, cond)).mean()) - base
|
|
assert abs(sym_push) > 1e-3 # mirror still pushes
|
|
assert abs(contra_push) < 1e-5 # collapsed axis -> no push
|
|
|
|
|
|
def test_contrastive_bad_pole_steers_away_from_bad(monkeypatch):
|
|
# bad_dir opposite to liked_dir: minus pole steps negative, plus pole positive, so the
|
|
# (eps_+ - eps_-) push is positive (toward liked, away from bad) — same sign as the
|
|
# symmetric case but driven by the real contrastive axis.
|
|
cond = torch.randn(1, 4, 16)
|
|
base = float(torch.full((1, 1, N), float(cond.float().mean())).mean())
|
|
contra, _ = _make_wrapper(2.0, monkeypatch, bad=-torch.ones(16))
|
|
push = float(contra(_apply_fn, _args(0.1, cond)).mean()) - base
|
|
assert push > 0
|