Integrate tiled upscale into Continuum second pass
This commit is contained in:
@@ -17,6 +17,16 @@ context-to-conditioning adaptation as **H3 Continuum Conditioning Bridge V3.5**.
|
||||
The separate `conditioning` input is optional for advanced legacy graphs; it is
|
||||
not needed in the normal Continuum wiring.
|
||||
|
||||
The recommended production wiring is now the **H3 Continuum Second Pass V3.5**
|
||||
node with `tiled_upscale = true`. That node receives the main sampler's video
|
||||
and audio latent lists plus its `refine_context` and `assembly_plan`, rebuilds a
|
||||
complete AV latent for each physical group, and runs the native tiled upscale
|
||||
one group at a time. It returns the upscaled video latent list while preserving
|
||||
the first-pass audio latent list. Connect the latent-upscale and spatial-split
|
||||
parameter nodes to its optional tiled inputs; its sampler, scheduler, steps,
|
||||
denoise, CFG and noise seed are independent from the ordinary Second Pass
|
||||
settings.
|
||||
|
||||
The node is an adapter around ComfyUI's native MiniMax H3 tiled-upscale
|
||||
implementation; it does not include model weights or vendor the upstream
|
||||
implementation. Install this repository as a normal custom-node folder and
|
||||
|
||||
+94
-21
@@ -29,6 +29,7 @@ from .refine_schedule import (
|
||||
serializable_schedule_contract,
|
||||
)
|
||||
from .refine_sampling import sample_refine_chunk
|
||||
from .tiled_latent_upscale_nodes import _call_native
|
||||
|
||||
|
||||
class SecondPassContractError(ValueError):
|
||||
@@ -457,6 +458,7 @@ def run_second_pass_groups(
|
||||
refine_schedule: RefineSchedule | None = None,
|
||||
selected_group_indices: Sequence[int] | None = None,
|
||||
temporal_windows_by_group: Mapping[int, Any] | None = None,
|
||||
tiled_upscale_config: Mapping[str, Any] | None = None,
|
||||
) -> tuple[list[dict[str, Any]], list[Any], dict[str, Any], str]:
|
||||
"""Refine complete physical groups while preserving first-pass audio exactly."""
|
||||
|
||||
@@ -501,6 +503,7 @@ def run_second_pass_groups(
|
||||
context_frames = int(detail["context_frames"])
|
||||
physical_seed = derive_refine_seed(int(refine_seed), group_index)
|
||||
group_seeds.append(physical_seed)
|
||||
upscale_seed = None
|
||||
conditioning_sources.append(conditioning_source)
|
||||
route_record = detail.get("reference_routing_v1")
|
||||
inheritance = (
|
||||
@@ -514,27 +517,69 @@ def run_second_pass_groups(
|
||||
"contract_sha256": route_record.get("sha256") if route_record is not None else None,
|
||||
})
|
||||
nested_latent = latent_builder(video_samples, audio_samples)
|
||||
sample_arguments = dict(
|
||||
model=chunk_model,
|
||||
conditioning=conditioning,
|
||||
latent=nested_latent,
|
||||
sampler=sampler,
|
||||
sigmas=effective_sigmas,
|
||||
seed=physical_seed,
|
||||
enable_preview=bool(enable_preview),
|
||||
)
|
||||
if temporal_windows_by_group is not None:
|
||||
temporal_window = temporal_windows_by_group.get(group_index)
|
||||
if temporal_window is None:
|
||||
raise SecondPassContractError(
|
||||
f"selected physical group {group_index + 1} has no temporal window"
|
||||
if tiled_upscale_config is None:
|
||||
sample_arguments = dict(
|
||||
model=chunk_model,
|
||||
conditioning=conditioning,
|
||||
latent=nested_latent,
|
||||
sampler=sampler,
|
||||
sigmas=effective_sigmas,
|
||||
seed=physical_seed,
|
||||
enable_preview=bool(enable_preview),
|
||||
)
|
||||
if temporal_windows_by_group is not None:
|
||||
temporal_window = temporal_windows_by_group.get(group_index)
|
||||
if temporal_window is None:
|
||||
raise SecondPassContractError(
|
||||
f"selected physical group {group_index + 1} has no temporal window"
|
||||
)
|
||||
sample_arguments["temporal_window"] = temporal_window
|
||||
sampled = sample_fn(**sample_arguments)
|
||||
else:
|
||||
config = tiled_upscale_config
|
||||
native = config["native"]
|
||||
upscale_seed = derive_refine_seed(
|
||||
int(config.get("seed_base", physical_seed)), group_index
|
||||
)
|
||||
noise = config["random_noise"]().get_noise(upscale_seed)
|
||||
upscale_sampler = config["sampler"]
|
||||
upscale_sigmas = config.get("sigmas")
|
||||
if upscale_sigmas is None:
|
||||
upscale_sigmas = config["scheduler"]().get_sigmas(
|
||||
chunk_model,
|
||||
config["scheduler_name"],
|
||||
config["steps"],
|
||||
config["denoise"],
|
||||
)
|
||||
sample_arguments["temporal_window"] = temporal_window
|
||||
sampled = sample_fn(**sample_arguments)
|
||||
native_kwargs = {
|
||||
"model": chunk_model,
|
||||
"conditioning": conditioning,
|
||||
"latent": nested_latent,
|
||||
"noise": noise,
|
||||
"sampler": upscale_sampler,
|
||||
"sigmas": upscale_sigmas,
|
||||
"cfg": config["cfg"],
|
||||
"latent_upscale_param": config["latent_upscale_param"],
|
||||
"temporal_split_param": config.get("temporal_split_param"),
|
||||
"spatial_split_param": config["spatial_split_param"],
|
||||
}
|
||||
if config.get("negative") is not None:
|
||||
native_kwargs["negative"] = config["negative"]
|
||||
native_result = _call_native(native, native_kwargs)
|
||||
if not isinstance(native_result, (tuple, list)) or not native_result:
|
||||
raise SecondPassContractError(
|
||||
f"tiled upscale returned no output for physical group {group_index + 1}"
|
||||
)
|
||||
sampled = native_result[0]
|
||||
refined_video, _temporary_audio = stream_extractor(sampled)
|
||||
if tuple(refined_video.shape) != tuple(video_samples.shape):
|
||||
if tiled_upscale_config is None:
|
||||
if tuple(refined_video.shape) != tuple(video_samples.shape):
|
||||
raise SecondPassContractError(
|
||||
f"refined video group {group_index} changed B/C/T/H/W geometry"
|
||||
)
|
||||
elif tuple(refined_video.shape[:3]) != tuple(video_samples.shape[:3]):
|
||||
raise SecondPassContractError(
|
||||
f"refined video group {group_index} changed B/C/T/H/W geometry"
|
||||
f"tiled upscale group {group_index} changed B/C/T geometry"
|
||||
)
|
||||
if not bool(torch.isfinite(refined_video.float()).all().item()):
|
||||
raise SecondPassContractError(
|
||||
@@ -554,8 +599,11 @@ def run_second_pass_groups(
|
||||
f"reference_inheritance={inheritance}, "
|
||||
f"physical_clip_index={physical_clip_index}, "
|
||||
f"context_frames={context_frames}, "
|
||||
f"refine_seed={physical_seed}, shape={tuple(refined_video.shape)}, "
|
||||
"sampling_passes=1, temporary_audio_discarded=true"
|
||||
f"refine_seed={physical_seed}, "
|
||||
f"upscale_seed={upscale_seed if upscale_seed is not None else 'n/a'}, "
|
||||
f"shape={tuple(refined_video.shape)}, "
|
||||
f"sampling_passes=1, tiled_upscale={tiled_upscale_config is not None}, "
|
||||
"temporary_audio_discarded=true"
|
||||
)
|
||||
if adaptation_stats is not None:
|
||||
group_report_lines.append(
|
||||
@@ -619,11 +667,36 @@ def run_second_pass_groups(
|
||||
|
||||
passthrough_audio = passthrough_audio_latents(audio_latents)
|
||||
updated_plan = prepared["updated_assembly_plan"]
|
||||
if tiled_upscale_config is not None:
|
||||
if selection_active:
|
||||
raise SecondPassContractError(
|
||||
"tiled upscale mode requires all physical groups; selective tiled "
|
||||
"upscale would produce mixed output geometry"
|
||||
)
|
||||
output_samples = _latent_samples(
|
||||
refined_videos[0], name="tiled refined video group 0"
|
||||
)
|
||||
target_height = int(output_samples.shape[-2]) * 16
|
||||
target_width = int(output_samples.shape[-1]) * 16
|
||||
updated_plan = copy.deepcopy(updated_plan)
|
||||
updated_plan["height"] = target_height
|
||||
updated_plan["width"] = target_width
|
||||
updated_plan["second_pass_contract"]["target_height"] = target_height
|
||||
updated_plan["second_pass_contract"]["target_width"] = target_width
|
||||
for group in updated_plan["second_pass_contract"]["physical_groups"]:
|
||||
group["target_latent_h"] = int(output_samples.shape[-2])
|
||||
group["target_latent_w"] = int(output_samples.shape[-1])
|
||||
group["target_height"] = target_height
|
||||
group["target_width"] = target_width
|
||||
contract = updated_plan["second_pass_contract"]
|
||||
if conditioning_sources and all(
|
||||
source == "refine_context" for source in conditioning_sources
|
||||
):
|
||||
contract["execution"] = "context_aware_physical_groups_audio_locked_v3"
|
||||
contract["execution"] = (
|
||||
"context_aware_physical_groups_tiled_upscale_audio_locked_v1"
|
||||
if tiled_upscale_config is not None
|
||||
else "context_aware_physical_groups_audio_locked_v3"
|
||||
)
|
||||
elif any(source == "refine_context" for source in conditioning_sources):
|
||||
contract["execution"] = "mixed_context_physical_groups_audio_locked_v3"
|
||||
else:
|
||||
|
||||
@@ -49,6 +49,22 @@ def _single_list_input(name: str, value: Any) -> Any:
|
||||
return value[0]
|
||||
|
||||
|
||||
def _sampler_options():
|
||||
try:
|
||||
import comfy.samplers
|
||||
return list(comfy.samplers.KSampler.SAMPLERS)
|
||||
except Exception:
|
||||
return ["euler", "euler_ancestral", "res_multistep", "sa_solver"]
|
||||
|
||||
|
||||
def _scheduler_options():
|
||||
try:
|
||||
import comfy.samplers
|
||||
return list(comfy.samplers.KSampler.SCHEDULERS)
|
||||
except Exception:
|
||||
return ["simple", "beta", "sgm_uniform", "karras", "exponential"]
|
||||
|
||||
|
||||
class H3ContinuumSecondPassV35:
|
||||
"""Low-sigma refinement over complete physical decode groups."""
|
||||
|
||||
@@ -83,10 +99,28 @@ class H3ContinuumSecondPassV35:
|
||||
"tooltip": "Base seed independently derived once per physical group.",
|
||||
},
|
||||
),
|
||||
"tiled_upscale": (
|
||||
"BOOLEAN",
|
||||
{
|
||||
"default": False,
|
||||
"tooltip": "Use native tiled H3 latent upscale instead of the low-sigma refine pass.",
|
||||
},
|
||||
),
|
||||
},
|
||||
"optional": {
|
||||
"refine_context": ("H3_CONTINUUM_REFINE_CONTEXT",),
|
||||
"video_vae": ("VAE",),
|
||||
"latent_upscale_param": ("H3_UPSCALE_PARAM",),
|
||||
"spatial_split_param": ("H3_SPATIAL_PARAM",),
|
||||
"temporal_split_param": ("H3_TEMPORAL_PARAM",),
|
||||
"upscale_sampler": (_sampler_options(), {"default": "euler_ancestral"}),
|
||||
"upscale_scheduler": (_scheduler_options(), {"default": "simple"}),
|
||||
"upscale_steps": ("INT", {"default": 3, "min": 1, "max": 50}),
|
||||
"upscale_denoise": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.01}),
|
||||
"upscale_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 20.0, "step": 0.1}),
|
||||
"upscale_noise_seed": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFFFFFFFFFFFF, "control_after_generate": True}),
|
||||
"upscale_negative": ("CONDITIONING",),
|
||||
"upscale_sigmas": ("SIGMAS",),
|
||||
},
|
||||
}
|
||||
|
||||
@@ -115,7 +149,66 @@ class H3ContinuumSecondPassV35:
|
||||
refine_context=None,
|
||||
video_vae=None,
|
||||
refine_schedule=None,
|
||||
tiled_upscale=False,
|
||||
latent_upscale_param=None,
|
||||
spatial_split_param=None,
|
||||
temporal_split_param=None,
|
||||
upscale_sampler="euler_ancestral",
|
||||
upscale_scheduler="simple",
|
||||
upscale_steps=3,
|
||||
upscale_denoise=0.2,
|
||||
upscale_cfg=1.0,
|
||||
upscale_noise_seed=0,
|
||||
upscale_negative=None,
|
||||
upscale_sigmas=None,
|
||||
):
|
||||
tiled_config = None
|
||||
if bool(_single_list_input("tiled_upscale", tiled_upscale)):
|
||||
from .tiled_latent_upscale_nodes import _native_node
|
||||
|
||||
required_native = {
|
||||
name: _native_node(name)
|
||||
for name in (
|
||||
"MMH3UltimateUpscale",
|
||||
"RandomNoise",
|
||||
"KSamplerSelect",
|
||||
"BasicScheduler",
|
||||
)
|
||||
}
|
||||
missing = [name for name, value in required_native.items() if value is None]
|
||||
if missing:
|
||||
raise RuntimeError("Missing native tiled-upscale nodes: " + ", ".join(missing))
|
||||
if latent_upscale_param is None or spatial_split_param is None:
|
||||
raise ValueError(
|
||||
"Tiled upscale mode requires latent_upscale_param and spatial_split_param"
|
||||
)
|
||||
tiled_config = {
|
||||
"native": required_native["MMH3UltimateUpscale"],
|
||||
"random_noise": required_native["RandomNoise"],
|
||||
"sampler": required_native["KSamplerSelect"]().get_sampler(
|
||||
str(_single_list_input("upscale_sampler", upscale_sampler))
|
||||
),
|
||||
"scheduler": required_native["BasicScheduler"],
|
||||
"scheduler_name": str(_single_list_input("upscale_scheduler", upscale_scheduler)),
|
||||
"steps": int(_single_list_input("upscale_steps", upscale_steps)),
|
||||
"denoise": float(_single_list_input("upscale_denoise", upscale_denoise)),
|
||||
"cfg": float(_single_list_input("upscale_cfg", upscale_cfg)),
|
||||
"seed_base": int(_single_list_input("upscale_noise_seed", upscale_noise_seed)),
|
||||
"latent_upscale_param": _single_list_input("latent_upscale_param", latent_upscale_param),
|
||||
"spatial_split_param": _single_list_input("spatial_split_param", spatial_split_param),
|
||||
"temporal_split_param": (
|
||||
None if temporal_split_param is None
|
||||
else _single_list_input("temporal_split_param", temporal_split_param)
|
||||
),
|
||||
"negative": (
|
||||
None if upscale_negative is None
|
||||
else _single_list_input("upscale_negative", upscale_negative)
|
||||
),
|
||||
"sigmas": (
|
||||
None if upscale_sigmas is None
|
||||
else _single_list_input("upscale_sigmas", upscale_sigmas)
|
||||
),
|
||||
}
|
||||
return run_second_pass_groups(
|
||||
model=_single_list_input("model", model),
|
||||
clip=_single_list_input("clip", clip),
|
||||
@@ -136,6 +229,7 @@ class H3ContinuumSecondPassV35:
|
||||
else _single_list_input("video_vae", video_vae)
|
||||
),
|
||||
refine_schedule=refine_schedule,
|
||||
tiled_upscale_config=tiled_config,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user