nodes/olivv-cs-ComfyUI-FunPack
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: H3 picture found in the packed model output; latent steering was off on every real run

Browse Source 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
digital-garbage
2026-09-27 19:01:13 +03:00
co-authored by Claude Opus 5.5
parent 169e8a6a1e
commit 0bb7e3a01a
12 changed files with 124 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sampling/dynashift/__init__.py
+1 -1
View File
@@ -188,7 +188,7 @@ def install(patcher, values, key):
out = executor(x, t, *args, **kwargs)
named = streams.model_args(args, kwargs)
to = named.get("transformer_options")
split = streams.video_of(out)
split = streams.video_of(out, named)
if split is None:
_say("off this run: could not find the picture in this model's latent")
return out
modules/sampling/dynashift/tests/test_dynashift.py
+7 -7
View File
@@ -51,7 +51,6 @@ def test_nothing_banked_at_this_resolution_changes_nothing():
def test_it_banks_the_last_step_and_steers_once_rated(tiny_h3, monkeypatch):
from comfy.nested_tensor import NestedTensor
from comfy.patcher_extension import WrappersMP
from modules.sampling.modifiers.nodes import FunPackLoadModifiers
from modules.system.taste import store
@@ -66,12 +65,13 @@ def test_it_banks_the_last_step_and_steers_once_rated(tiny_h3, monkeypatch):
torch.manual_seed(1)
bad = torch.randn(1, 4, 3, 8, 8)
audio = torch.randn(1, 8, 5)
x0 = NestedTensor([bad, audio])
from conftest import packed_av, unpacked
x0, shapes = packed_av(bad, audio)
for i, rating in enumerate(["disliked", "disliked"]):
monkeypatch.setattr(store, "current_prompt_id", lambda i=i: f"run-{i}")
wrap, outer, status = load()
for o in outer:
o(lambda: wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3)))
o(lambda: wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3), latent_shapes=shapes))
assert store.rate(f"run-{i}", rating)["recorded"] == ["dynashift"]
wrap, outer, _status = load()
@@ -80,8 +80,8 @@ def test_it_banks_the_last_step_and_steers_once_rated(tiny_h3, monkeypatch):
for o in outer:
o(lambda: None) # a run starts: the bank is read
assert any("2 disliked, 0 liked banked" in e["message"] for e in log.history())
early = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(1))
early = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(1), latent_shapes=shapes)
assert early is x0 # first half: untouched
late = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3))
assert not torch.allclose(late.tensors[0], bad)
assert late.tensors[1] is audio # sound untouched
late = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3), latent_shapes=shapes)
assert not torch.allclose(unpacked(late, shapes)[0], bad)
assert torch.equal(unpacked(late, shapes)[1], audio) # sound untouched
modules/sampling/output_guidance/__init__.py
+1 -1
View File
@@ -74,7 +74,7 @@ def install(patcher, values, key):
out = executor(x, t, *args, **kwargs)
named = streams.model_args(args, kwargs)
to = named.get("transformer_options")
split = streams.video_of(out)
split = streams.video_of(out, named)
if split is None:
_say("off this run: could not find the picture in this model's latent")
return out
modules/sampling/output_guidance/tests/test_output_guidance.py
+9 -8
View File
@@ -42,9 +42,9 @@ def _teach(kind, names):
def _x0():
from comfy.nested_tensor import NestedTensor
from conftest import packed_av
torch.manual_seed(3)
return NestedTensor([torch.randn(1, 4, 3, 8, 8), torch.randn(1, 8, 5)])
return packed_av(torch.randn(1, 4, 3, 8, 8), torch.randn(1, 8, 5))
def _start(outer):
@@ -55,16 +55,17 @@ def _start(outer):
def test_late_guidance_banks_the_final_picture_and_steers_only_late(tiny_h3, monkeypatch):
from modules.system.taste import store
monkeypatch.setattr(store, "current_prompt_id", lambda: "run-1")
x0 = _x0()
x0, shapes = _x0()
wrap, outer = _load(tiny_h3, "output_guidance")
for o in outer:
o(lambda: wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3)))
o(lambda: wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3), latent_shapes=shapes))
assert store.rate("run-1", "liked")["recorded"] == ["x0_final"]
_teach("x0_final", ["final"])
wrap, outer = _load(tiny_h3, "output_guidance")
_start(outer)
assert wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(1)) is x0
late = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3))
assert not torch.allclose(late.tensors[0], x0.tensors[0])
assert late.tensors[1] is x0.tensors[1]
from conftest import unpacked
assert wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(1), latent_shapes=shapes) is x0
late = wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(3), latent_shapes=shapes)
assert not torch.allclose(unpacked(late, shapes)[0], unpacked(x0, shapes)[0])
assert torch.equal(unpacked(late, shapes)[1], unpacked(x0, shapes)[1])
modules/sampling/score_slider/__init__.py
+3 -3
View File
@@ -120,12 +120,12 @@ def install(patcher, values, key):
return base
size = torch.linalg.vector_norm(c[:, words], dim=-1, dtype=torch.float32).mean()
step = (d.to(c.device) * STEP * size).to(c.dtype) * words.view(1, -1, 1)
split = streams.video_of(base)
split = streams.video_of(base, named)
if split is None:
_say("off this run: could not find the picture in this model's latent")
return base
plus = streams.video_of(executor(x, t, **{**named, "c_crossattn": c + step}))
minus = streams.video_of(executor(x, t, **{**named, "c_crossattn": c - step}))
plus = streams.video_of(executor(x, t, **{**named, "c_crossattn": c + step}), named)
minus = streams.video_of(executor(x, t, **{**named, "c_crossattn": c - step}), named)
video, rebuild = split
return rebuild(video + (plus[0] - minus[0]) * amount)
modules/sampling/score_slider/tests/test_score_slider.py
+8 -7
View File
@@ -44,7 +44,7 @@ def test_similar_prompts_pick_their_own_direction():
def test_late_steps_combine_three_passes_on_the_picture_words_only(tiny_h3):
from comfy.nested_tensor import NestedTensor
from conftest import packed_av, unpacked
from comfy.patcher_extension import WrappersMP
from modules.sampling.modifiers.nodes import FunPackLoadModifiers
from modules.system.taste import store
@@ -68,19 +68,20 @@ def test_late_steps_combine_three_passes_on_the_picture_words_only(tiny_h3):
def executor(x, t, c_concat=None, c_crossattn=None, control=None, transformer_options=None, **kw):
seen.append(c_crossattn.clone())
return NestedTensor([torch.full((1, 2, 1, 2, 2), float(c_crossattn[0, :, 0].sum())), audio])
return packed_av(torch.full((1, 2, 1, 2, 2), float(c_crossattn[0, :, 0].sum())), audio)[0]
c = torch.ones(1, 4, dim)
tags = torch.tensor([1, 1, 0, 1]) # row 2: a reference image's token
x0 = executor(None, None, c_crossattn=c)
shapes = packed_av(torch.zeros(1, 2, 1, 2, 2), audio)[1]
seen.clear()
early = wrap(executor, x0, None, None, c, None, _to(0), minimax_payload={"text_token_tags": tags})
assert len(seen) == 1 and torch.equal(early.tensors[0], x0.tensors[0])
early = wrap(executor, x0, None, None, c, None, _to(0), minimax_payload={"text_token_tags": tags}, latent_shapes=shapes)
assert len(seen) == 1 and torch.equal(early, x0)
seen.clear()
late = wrap(executor, x0, None, None, c, None, _to(3), minimax_payload={"text_token_tags": tags})
late = wrap(executor, x0, None, None, c, None, _to(3), minimax_payload={"text_token_tags": tags}, latent_shapes=shapes)
assert len(seen) == 3
plus = seen[1]
assert torch.equal(plus[0, 2], c[0, 2]) and not torch.equal(plus[0, 0], c[0, 0])
assert not torch.equal(late.tensors[0], x0.tensors[0])
assert late.tensors[1] is audio
assert not torch.equal(unpacked(late, shapes)[0], unpacked(x0, shapes)[0])
assert torch.equal(unpacked(late, shapes)[1], audio)
modules/sampling/trajectory_guidance/__init__.py
+1 -1
View File
@@ -80,7 +80,7 @@ def install(patcher, values, key):
out = executor(x, t, *args, **kwargs)
named = streams.model_args(args, kwargs)
q = quarter(named.get("transformer_options"))
split = streams.video_of(out)
split = streams.video_of(out, named)
if q is None or split is None:
_say("off this run: could not place this step on the schedule or find the picture")
return out
modules/sampling/trajectory_guidance/tests/test_trajectory_guidance.py
+6 -5
View File
@@ -42,9 +42,9 @@ def _teach(kind, names):
def _x0():
from comfy.nested_tensor import NestedTensor
from conftest import packed_av
torch.manual_seed(3)
return NestedTensor([torch.randn(1, 4, 3, 8, 8), torch.randn(1, 8, 5)])
return packed_av(torch.randn(1, 4, 3, 8, 8), torch.randn(1, 8, 5))
def _start(outer):
@@ -56,17 +56,18 @@ def test_early_guidance_steers_in_the_first_quarter_and_banks_every_quarter(tiny
from modules.system.taste import store
_teach("x0_quarters", ["q0", "q1", "q2", "q3"])
monkeypatch.setattr(store, "current_prompt_id", lambda: "run-2")
x0 = _x0()
x0, shapes = _x0()
wrap, outer = _load(tiny_h3, "trajectory_guidance")
def run():
return [wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(i)) for i in range(4)]
return [wrap(lambda *a, **k: x0, x0, None, None, None, None, _to(i), latent_shapes=shapes) for i in range(4)]
results = {}
for o in outer:
o(lambda: results.setdefault("steps", run()))
first = results["steps"][0]
assert not torch.allclose(first.tensors[0], x0.tensors[0])
assert not torch.allclose(first[..., :4 * 3 * 8 * 8], x0[..., :4 * 3 * 8 * 8])
assert torch.equal(first[..., 4 * 3 * 8 * 8:], x0[..., 4 * 3 * 8 * 8:])
pending = torch.load(store.ROOT / "fox" / "x0_quarters.pending.pt")["rows"]
assert sorted(pending) == ["q0", "q1", "q2", "q3"]