prune failed churn-experiment hires modes (frozen/faint/converge); dropdown = subtle/medium/strong/spatial with tournament-informed tooltip

nodes.py

@@ -1209,17 +1209,17 @@ class JoyEcho_Generate:
                 }),
                 "hires_denoise": (["subtle (1 step)", "medium (2 steps)",
                                    "strong (tenstrip 4-step)",
-                                   "subtle (frozen noise, experimental)",
-                                   "faint (1 step from 0.30)",
-                                   "subtle-converge (3 shallow steps)",
                                    "spatial (LTX latent upsampler, no churn)"], {
                     "default": "subtle (1 step)",
-                    "tooltip": "How deep the refine re-noises. subtle = from sigma 0.42, one "
-                               "denoise step: adds texture, very faithful to the base. medium "
-                               "= from 0.725, two steps: more synthesis, more drift risk. "
-                               "strong = tenstrip's published upscale ladder "
-                               "(0.92/0.725/0.42, three steps): deepest synthesis - only "
-                               "sensible now that all windows share one per-shot noise field.",
+                    "tooltip": "How the hires pass works. The three refine modes re-noise and "
+                               "re-denoise at target res - they SYNTHESIZE real detail but "
+                               "reshuffle fine texture slightly per frame (subtle = sigma 0.42 "
+                               "one step, most faithful; medium = 0.725 two steps; strong = "
+                               "tenstrip's 0.92/0.725/0.42 ladder, deepest). spatial = the LTX "
+                               "latent upsampler on the shot's own latents: deterministic, "
+                               "zero temporal churn, fixed 1.5x (hires_factor just enables "
+                               "it) - use heights whose /32 is EVEN (768, not 736) or it "
+                               "smears one edge.",
                 }),
             },
         }
@@ -1966,16 +1966,10 @@ class JoyEcho_Generate:
         SIGMA_TAILS = {"subtle": [0.421875, 0.0], "medium": [0.725, 0.421875, 0.0],
                        # tenstrip's published UPSCALE ladder (HF, ~Jul 1) -
                        # purpose-built for a refine pass, unlike the generation
-                       # tails above.
-                       "strong": [0.92, 0.725, 0.421875, 0.0],
-                       # churn experiments (2026-07-23): depth worsens per-frame
-                       # detail churn monotonically, so these go SHALLOWER.
-                       # faint = less noise in, less residue out.
-                       # subtle-converge = extra steps confined to the shallow
-                       # range, with re-corruption REUSING the initial noise
-                       # field (randomness enters once; steps only converge).
-                       "faint": [0.30, 0.0],
-                       "subtle-converge": [0.421875, 0.28, 0.14, 0.0]}
+                       # tails above. (Shallower/frozen/reused-noise variants
+                       # were tested 2026-07-23 and removed: none reduced the
+                       # per-frame detail churn - see project notes.)
+                       "strong": [0.92, 0.725, 0.421875, 0.0]}
         sigmas = SIGMA_TAILS.get(str(mode).split()[0].lower(), SIGMA_TAILS["subtle"])
         base_h, base_w = int(frames_list[0].shape[1]), int(frames_list[0].shape[2])
         th = max(32, int(round(base_h * factor / 32.0)) * 32)
@@ -2097,17 +2091,7 @@ class JoyEcho_Generate:
                     # multiples of (WIN-OVL)=24 pixel frames = 3 latent frames,
                     # so overlapping windows land on the same global indices.
                     ls = min(int(round(s / 8.0)), shot_fields["v"][0].shape[1] - Fl)
-                    # "frozen noise" (experimental): one spatial noise pattern
-                    # repeated for every VIDEO latent frame, so the single-step
-                    # residue becomes a static grain plate instead of per-frame
-                    # detail churn. Audio noise stays per-frame (refined audio
-                    # is discarded anyway).
-                    _frozen = "frozen" in str(mode).lower()
-                    if _frozen:
-                        nv = shot_fields["v"][0][:, :1].repeat(1, Fl, 1, 1, 1).to(
-                            device=device, dtype=_dt)
-                    else:
-                        nv = shot_fields["v"][0][:, ls:ls + Fl].to(device=device, dtype=_dt)
+                    nv = shot_fields["v"][0][:, ls:ls + Fl].to(device=device, dtype=_dt)
                     na = shot_fields["a"][0][:, a0:a0 + Fa_w].to(device=device, dtype=_dt)
                     s0 = torch.full((1, Fl), sigmas[0])
                     sa0 = torch.full((1, Fa_w), sigmas[0])
@@ -2126,14 +2110,10 @@ class JoyEcho_Generate:
                         if nxt > 0:
                             # re-corruption noise from the step's shared field,
                             # same global slice - never randn_like (that was
-                            # fresh unseeded noise per window per step in v1).
-                            # converge mode reuses the INITIAL field every step:
-                            # randomness enters once, further steps only pull
-                            # structure toward a consistent answer.
-                            _ev = 0 if "converge" in str(mode).lower() else i_s + 1
-                            fnv = shot_fields["v"][_ev][:, ls:ls + Fl].to(
+                            # fresh unseeded noise per window per step in v1)
+                            fnv = shot_fields["v"][i_s + 1][:, ls:ls + Fl].to(
                                 device=device, dtype=_dt)
-                            fna = shot_fields["a"][_ev][:, a0:a0 + Fa_w].to(
+                            fna = shot_fields["a"][i_s + 1][:, a0:a0 + Fa_w].to(
                                 device=device, dtype=_dt)
                             nvs = nxt * torch.ones((1, Fl), device=device)
                             nas = nxt * torch.ones((1, Fa_w), device=device)