Harden H3 detail pass conditioning

2026-08-28 09:37:22 +00:00
parent 5255c174fd
commit af226987d7
2 changed files with 77 additions and 7 deletions
@@ -4031,6 +4031,21 @@ def _copy_sample_latent(out_latent):
        return None


+def _latent_with_replaced_samples(template_latent, sampled_latent):
+    """Reuse the original latent payload, but swap in freshly sampled tensors."""
+    if not isinstance(template_latent, dict):
+        return sampled_latent
+    out = dict(template_latent)
+    if isinstance(sampled_latent, dict):
+        for key, value in sampled_latent.items():
+            if key != "samples" and key not in out:
+                out[key] = value
+        if "samples" in sampled_latent:
+            out["samples"] = sampled_latent["samples"]
+            return out
+    return sampled_latent
+
+
 def _video_only_refined_latent(base_latent, refined_latent):
    """Keep the refined video latent, but preserve the original audio latent."""
    base = base_latent.get("samples") if isinstance(base_latent, dict) else None
@@ -4049,6 +4064,16 @@ def _video_only_refined_latent(base_latent, refined_latent):
    return refined_latent


+def _coerce_bool_flag(value):
+    if isinstance(value, str):
+        text = value.strip().lower()
+        if text in ("", "0", "false", "no", "off", "none", "null"):
+            return False
+        if text in ("1", "true", "yes", "on"):
+            return True
+    return bool(value)
+
+
 # --- ref2va reference conditioning ----------------------------------------
 # H3's reference pipeline encodes a reference image at up to a 2048 short edge.
 # Reference rows ride through EVERY sampling step, so this is also the setting
@@ -6148,7 +6173,7 @@ class H3LongVideos:
                               "building a separate graph."}),
                "detail_sampler_name": (comfy.samplers.KSampler.SAMPLERS, {"default": "euler",
                    "tooltip": "Sampler used for the optional refinement pass."}),
-                "detail_scheduler": (comfy.samplers.KSampler.SCHEDULERS, {"default": "karras",
+                "detail_scheduler": (comfy.samplers.KSampler.SCHEDULERS, {"default": "beta",
                    "tooltip": "Scheduler used for the optional refinement pass."}),
                "detail_steps": ("INT", {"default": 8, "min": 1, "max": 200,
                    "tooltip": "Steps for the optional refinement pass."}),
@@ -6178,7 +6203,7 @@ class H3LongVideos:
    def _render(self, model, clip, vae, audio_vae, negative, prompt, w, h, ln, fps, tiled, sa,
                handoff, decode_tile_frames=0, decode_tile_size=0,
                refs=None, ref_image_size="match", ref_noise_aug=None, silent=False,
-                detail_pass=False, detail_sampler_name="euler", detail_scheduler="karras",
+                detail_pass=False, detail_sampler_name="euler", detail_scheduler="beta",
                detail_steps=8, detail_denoise=0.4):
        positive, latent = _build_shot_conditioning(clip, vae, prompt, w, h, ln, fps, handoff,
                                                    ref_images=refs, ref_image_size=ref_image_size,
@@ -6200,11 +6225,13 @@ class H3LongVideos:
                e._h3_stage = "sampling"
            raise
        refined_out = out
+        detail_pass = _coerce_bool_flag(detail_pass)
        if detail_pass:
+            detail_latent = _latent_with_replaced_samples(latent, out)
            try:
                (refined_out,) = nodes.common_ksampler(
                    model, seed, int(detail_steps), cfg, detail_sampler_name, detail_scheduler,
-                    positive, negative, out, denoise=float(detail_denoise))
+                    positive, negative, detail_latent, denoise=float(detail_denoise))
            except Exception as e:
                if _is_oom(e):
                    e._h3_stage = "sampling"
@@ -6254,7 +6281,7 @@ class H3LongVideos:
            ref_image_5=None, ref_image_6=None, ref_image_7=None, ref_image_8=None,
            ref_image_9=None,
            ref_mode="auto ref2v", ref_image_size="match", ref_noise_aug=0.95,
-            detail_pass=False, detail_sampler_name="euler", detail_scheduler="karras",
+            detail_pass=False, detail_sampler_name="euler", detail_scheduler="beta",
            detail_steps=8, detail_denoise=0.4,
            graph=None, node_id=None):

@@ -226,15 +226,16 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
                handoff=None,
                detail_pass=True,
                detail_sampler_name="euler",
-                detail_scheduler="karras",
+                detail_scheduler="beta",
                detail_steps=5,
                detail_denoise=0.4,
            )

            self.assertEqual(len(calls), 2)
-            self.assertIs(calls[1][0][8], first_out)
+            self.assertIsNot(calls[1][0][8], first_out)
+            self.assertIs(calls[1][0][8]["samples"], first_out["samples"])
            self.assertEqual(calls[1][0][4], "euler")
-            self.assertEqual(calls[1][0][5], "karras")
+            self.assertEqual(calls[1][0][5], "beta")
            self.assertAlmostEqual(calls[1][1]["denoise"], 0.4)
            self.assertEqual(result[1], first_out)
            self.assertEqual(result[2][0].name, "v2")
@@ -253,6 +254,48 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
            else:
                self.module.comfy.nested_tensor.NestedTensor = original_nested

+    def test_detail_pass_treats_falsey_strings_as_disabled(self):
+        calls = []
+        original_common_ksampler = self.module.nodes.common_ksampler
+        original_build = self.module._build_shot_conditioning
+        original_evict = self.module._evict_all_but
+        original_decode_video = self.module._decode_video
+        original_decode_audio = self.module._decode_audio
+        original_cleanup = self.module._deep_cleanup
+        try:
+            self.module.nodes.common_ksampler = lambda *args, **kwargs: (calls.append((args, kwargs)) or {"samples": "latent"},)
+            self.module._build_shot_conditioning = lambda *_args, **_kwargs: ("cond", {"samples": "base"})
+            self.module._evict_all_but = lambda *_args, **_kwargs: None
+            self.module._decode_video = lambda _vae, out_latent, *_args, **_kwargs: out_latent
+            self.module._decode_audio = lambda _vae, out_latent: out_latent
+            self.module._deep_cleanup = lambda: None
+
+            self.module.H3LongVideos()._render(
+                model=object(),
+                clip=object(),
+                vae=object(),
+                audio_vae=object(),
+                negative="negative",
+                prompt="beat",
+                w=128,
+                h=64,
+                ln=24,
+                fps=24,
+                tiled=False,
+                sa=(123, 20, 1.0, "res_multistep", "simple", 1.0),
+                handoff=None,
+                detail_pass="false",
+            )
+
+            self.assertEqual(len(calls), 1)
+        finally:
+            self.module.nodes.common_ksampler = original_common_ksampler
+            self.module._build_shot_conditioning = original_build
+            self.module._evict_all_but = original_evict
+            self.module._decode_video = original_decode_video
+            self.module._decode_audio = original_decode_audio
+            self.module._deep_cleanup = original_cleanup
+
    def test_distribute_generations_canonicalizes_per_shot_audio_and_anchor_directives(self):
        generations = self.module.distribute_generations(
            "",