Bugfixes (#11)

* Add MagicLatentAdapter node

Introduces MagicLatentAdapter for generating or adapting latents to match model expectations, including support for 5D (NCDHW) layouts and experimental FLUX/Qwen models. Adds mask alignment utilities to CADE nodes for robust mask blending, improves Gaussian blur to support 5D tensors, and updates preset defaults for improved results. Documentation and workflow updated to reflect new node and features.

* Bugfixes

1. Adjust edge mask gain logic and preset parameters. Fix depth path for CF (hard).
2. Memory leaking fix.
3. Quality improvement at all steps (Easy and SuperSimple)

mod/easy/mg_cade25_easy.py

@@ -1264,6 +1264,21 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
     # Spectral switching/adaptive low state
     spec_state = {"ema": None, "mode": "CFGZeroFD"}
 
+    # External reset hook to emulate fresh state per iteration without re-cloning the model
+    def _reset_state():
+        try:
+            prev_delta["t"] = None
+            sigma_seen["max"] = None
+            sigma_seen["min"] = None
+            spec_state["ema"] = None
+            spec_state["mode"] = "CFGZeroFD"
+        except Exception:
+            pass
+    try:
+        setattr(m, "mg_guidance_reset", _reset_state)
+    except Exception:
+        pass
+
     def cfg_func(args):
         cond = args["cond"]
         uncond = args["uncond"]
@@ -1748,11 +1763,8 @@ def _build_cf_edge_mask_from_step(image_bhwc: torch.Tensor, preset_step: str) ->
                 ed = (ed * g).clamp(0,1)
             except Exception:
                 pass
-        # Apply opacity + edge strength + blend factor (as extra gain for edges only)
-        total_gain = max(0.0, float(edge_alpha)) * max(0.0, float(edge_strength_mul)) * max(0.0, float(blend_factor))
-        # Keep at least alpha if blend_factor is set to 0 in presets
-        if total_gain == 0.0:
-            total_gain = max(0.0, float(edge_alpha))
+        # Apply opacity + edge strength (keep blend_factor only for ControlNet stage, not for mask amplitude)
+        total_gain = max(0.0, float(edge_alpha)) * max(0.0, float(edge_strength_mul))
         ed = (ed * total_gain).clamp(0,1)
         # Return BHWC single-channel
         return ed.unsqueeze(0).unsqueeze(-1)
@@ -2257,6 +2269,19 @@ class ComfyAdaptiveDetailEnhancer25:
                 clipseg_enable = False
                 # Depth gate cache for micro-detail injection (reuse per resolution)
                 depth_gate_cache = {"size": None, "mask": None}
+                # Prepare guided sampler once per node run to avoid cloning model each iteration
+                sampler_model = _wrap_model_with_guidance(
+                      model, guidance_mode, rescale_multiplier, momentum_beta, cfg_curve, perp_damp,
+                      use_zero_init=bool(use_zero_init), zero_init_steps=int(zero_init_steps),
+                      fdg_low=float(fdg_low), fdg_high=float(fdg_high), fdg_sigma=float(fdg_sigma),
+                      midfreq_enable=bool(midfreq_enable), midfreq_gain=float(midfreq_gain), midfreq_sigma_lo=float(midfreq_sigma_lo), midfreq_sigma_hi=float(midfreq_sigma_hi),
+                      ze_zero_steps=int(ze_res_zero_steps),
+                      ze_adaptive=bool(ze_adaptive), ze_r_switch_hi=float(ze_r_switch_hi), ze_r_switch_lo=float(ze_r_switch_lo),
+                      fdg_low_adaptive=bool(fdg_low_adaptive), fdg_low_min=float(fdg_low_min), fdg_low_max=float(fdg_low_max), fdg_ema_beta=float(fdg_ema_beta),
+                      use_local_mask=bool(onnx_local_guidance), mask_inside=float(onnx_mask_inside), mask_outside=float(onnx_mask_outside),
+                      mahiro_plus_enable=bool(muse_blend), mahiro_plus_strength=float(muse_blend_strength),
+                      eps_scale_enable=bool(eps_scale_enable), eps_scale=float(eps_scale)
+                  )
                 # check once more right before the loop starts
                 model_management.throw_exception_if_processing_interrupted()
                 for i in range(iterations):
@@ -2265,6 +2290,13 @@ class ComfyAdaptiveDetailEnhancer25:
                     if i % 2 == 0:
                         clear_gpu_and_ram_cache()
 
+                    # Reset guidance internal state so each iteration starts clean
+                    try:
+                        if hasattr(sampler_model, "mg_guidance_reset"):
+                            sampler_model.mg_guidance_reset()
+                    except Exception:
+                        pass
+
                     prev_samples = current_latent["samples"].clone().detach()
 
                     iter_seed = seed + i * 7777
@@ -2376,19 +2408,8 @@ class ComfyAdaptiveDetailEnhancer25:
                     except Exception:
                         pass
 
-                    # Guidance override via cfg_func when requested
-                    sampler_model = _wrap_model_with_guidance(
-                          model, guidance_mode, rescale_multiplier, momentum_beta, cfg_curve, perp_damp,
-                          use_zero_init=bool(use_zero_init), zero_init_steps=int(zero_init_steps),
-                          fdg_low=float(fdg_low), fdg_high=float(fdg_high), fdg_sigma=float(fdg_sigma),
-                          midfreq_enable=bool(midfreq_enable), midfreq_gain=float(midfreq_gain), midfreq_sigma_lo=float(midfreq_sigma_lo), midfreq_sigma_hi=float(midfreq_sigma_hi),
-                          ze_zero_steps=int(ze_res_zero_steps),
-                          ze_adaptive=bool(ze_adaptive), ze_r_switch_hi=float(ze_r_switch_hi), ze_r_switch_lo=float(ze_r_switch_lo),
-                          fdg_low_adaptive=bool(fdg_low_adaptive), fdg_low_min=float(fdg_low_min), fdg_low_max=float(fdg_low_max), fdg_ema_beta=float(fdg_ema_beta),
-                          use_local_mask=bool(onnx_local_guidance), mask_inside=float(onnx_mask_inside), mask_outside=float(onnx_mask_outside),
-                          mahiro_plus_enable=bool(muse_blend), mahiro_plus_strength=float(muse_blend_strength),
-                          eps_scale_enable=bool(eps_scale_enable), eps_scale=float(eps_scale)
-                      )
+                    # Sampler model prepared once above; reused across iterations (no-op here)
+                    sampler_model = sampler_model
 
                     # Local best-of-2 in ROI (hands/face), only on early iteration to limit overhead
                     try:
@@ -2755,6 +2776,12 @@ class ComfyAdaptiveDetailEnhancer25:
                 sa_patch.enable_crossattention_nag_patch(False)
             except Exception:
                 pass
+            # Turn off attention-entropy probe (AQClip Attn-mode) to avoid holding last maps
+            try:
+                if hasattr(sa_patch, "enable_attention_entropy_capture"):
+                    sa_patch.enable_attention_entropy_capture(False)
+            except Exception:
+                pass
             # Disable KV pruning as well (avoid leaking state)
             try:
                 if hasattr(sa_patch, "set_kv_prune"):

mod/hard/mg_cade25.py

@@ -150,7 +150,7 @@ def _clipseg_build_mask(image_bhwc: torch.Tensor,
                 cur = _encode_clip_image(image_bhwc, clip_vision, target_res=224)
                 dist = _clip_cosine_distance(cur, ref_embed)
                 if dist > float(ref_threshold):
-                    # up to +50% gain if сильно уехали
+                    # up to +50% gain if distance exceeds the reference threshold
                     gate = 1.0 + min(0.5, (dist - float(ref_threshold)) * 4.0)
                     m = m * gate
             except Exception:
@@ -919,6 +919,21 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
     # Spectral switching/adaptive low state
     spec_state = {"ema": None, "mode": "CFGZeroFD"}
 
+    # External reset hook to emulate fresh state per iteration without re-cloning the model
+    def _mg_guidance_reset():
+        try:
+            prev_delta["t"] = None
+            sigma_seen["max"] = None
+            sigma_seen["min"] = None
+            spec_state["ema"] = None
+            spec_state["mode"] = "CFGZeroFD"
+        except Exception:
+            pass
+    try:
+        setattr(m, "mg_guidance_reset", _mg_guidance_reset)
+    except Exception:
+        pass
+
     def cfg_func(args):
         cond = args["cond"]
         uncond = args["uncond"]
@@ -1580,14 +1595,14 @@ class ComfyAdaptiveDetailEnhancer25:
                     mask_last = pre_mask
                     om = pre_mask.movedim(-1, 1)
                     pre_area = float(om.mean().item())
-                    # One-time gentle damping from area
-                    try:
-                        if pre_area > 0.005:
-                            damp = 1.0 - min(0.10, 0.02 + pre_area * 0.08)
-                            current_denoise = max(0.10, current_denoise * damp)
-                            current_cfg = max(1.0, current_cfg * (1.0 - 0.005))
-                    except Exception:
-                        pass
+                    # One-time gentle damping from area (disabled to preserve outline precision)
+                    # try:
+                    #     if pre_area > 0.005:
+                    #         damp = 1.0 - min(0.10, 0.02 + pre_area * 0.08)
+                    #         current_denoise = max(0.10, current_denoise * damp)
+                    #         current_cfg = max(1.0, current_cfg * (1.0 - 0.005))
+                    # except Exception:
+                    #     pass
                 # Compact status
                 try:
                     clipseg_status = "on" if bool(clipseg_enable) and isinstance(clipseg_text, str) and clipseg_text.strip() != "" else "off"
@@ -1600,6 +1615,19 @@ class ComfyAdaptiveDetailEnhancer25:
                 clipseg_enable = False
                 # Depth gate cache for micro-detail injection (reuse per resolution)
                 depth_gate_cache = {"size": None, "mask": None}
+                # Prepare guided sampler once per node run to avoid cloning model each iteration
+                sampler_model = _wrap_model_with_guidance(
+                      model, guidance_mode, rescale_multiplier, momentum_beta, cfg_curve, perp_damp,
+                      use_zero_init=bool(use_zero_init), zero_init_steps=int(zero_init_steps),
+                      fdg_low=float(fdg_low), fdg_high=float(fdg_high), fdg_sigma=float(fdg_sigma),
+                      midfreq_enable=bool(False), midfreq_gain=float(0.0), midfreq_sigma_lo=float(0.8), midfreq_sigma_hi=float(2.0),
+                      ze_zero_steps=int(ze_res_zero_steps),
+                      ze_adaptive=bool(ze_adaptive), ze_r_switch_hi=float(ze_r_switch_hi), ze_r_switch_lo=float(ze_r_switch_lo),
+                      fdg_low_adaptive=bool(fdg_low_adaptive), fdg_low_min=float(fdg_low_min), fdg_low_max=float(fdg_low_max), fdg_ema_beta=float(fdg_ema_beta),
+                      use_local_mask=False, mask_inside=1.0, mask_outside=1.0,
+                      mahiro_plus_enable=bool(muse_blend), mahiro_plus_strength=float(muse_blend_strength),
+                      eps_scale_enable=bool(eps_scale_enable), eps_scale=float(eps_scale)
+                  )
                 # early interruption check before starting the loop
                 try:
                     model_management.throw_exception_if_processing_interrupted()
@@ -1613,6 +1641,13 @@ class ComfyAdaptiveDetailEnhancer25:
                     if i % 2 == 0:
                         clear_gpu_and_ram_cache()
 
+                    # Reset guidance internal state so each iteration starts clean
+                    try:
+                        if hasattr(sampler_model, "mg_guidance_reset"):
+                            sampler_model.mg_guidance_reset()
+                    except Exception:
+                        pass
+
                     prev_samples = current_latent["samples"].clone().detach()
 
                     iter_seed = seed + i * 7777
@@ -1661,18 +1696,8 @@ class ComfyAdaptiveDetailEnhancer25:
                     except Exception:
                         pass
 
-                    # Guidance override via cfg_func when requested
-                    sampler_model = _wrap_model_with_guidance(
-                          model, guidance_mode, rescale_multiplier, momentum_beta, cfg_curve, perp_damp,
-                          use_zero_init=bool(use_zero_init), zero_init_steps=int(zero_init_steps),
-                          fdg_low=float(fdg_low), fdg_high=float(fdg_high), fdg_sigma=float(fdg_sigma),
-                          midfreq_enable=bool(False), midfreq_gain=float(0.0), midfreq_sigma_lo=float(0.8), midfreq_sigma_hi=float(2.0),
-                          ze_zero_steps=int(ze_res_zero_steps),
-                          ze_adaptive=bool(ze_adaptive), ze_r_switch_hi=float(ze_r_switch_hi), ze_r_switch_lo=float(ze_r_switch_lo),
-                          fdg_low_adaptive=bool(fdg_low_adaptive), fdg_low_min=float(fdg_low_min), fdg_low_max=float(fdg_low_max), fdg_ema_beta=float(fdg_ema_beta),
-                          mahiro_plus_enable=bool(muse_blend), mahiro_plus_strength=float(muse_blend_strength),
-                          eps_scale_enable=bool(eps_scale_enable), eps_scale=float(eps_scale)
-                      )
+                    # Sampler model prepared once above; reuse it here (no-op assignment)
+                    sampler_model = sampler_model
 
                     if str(scheduler) == "MGHybrid":
                         try:
@@ -1921,6 +1946,12 @@ class ComfyAdaptiveDetailEnhancer25:
                 sa_patch.enable_crossattention_nag_patch(False)
             except Exception:
                 pass
+            # Turn off attention-entropy probe to avoid holding last maps
+            try:
+                if hasattr(sa_patch, "enable_attention_entropy_capture"):
+                    sa_patch.enable_attention_entropy_capture(False)
+            except Exception:
+                pass
             try:
                 sa_patch.CURRENT_PV_ACCUM = prev_accum
             except Exception:

mod/hard/mg_controlfusion.py

@@ -30,6 +30,47 @@ def _try_init_depth_anything(model_path: str):
     if _DEPTH_INIT:
         return _DEPTH_MODEL is not None
     _DEPTH_INIT = True
+    # Resolve model path: allow 'auto' or directory and prefer vitl>vitb>vits>vitg
+    try:
+        def _prefer_order(paths):
+            order = ["vitl", "vitb", "vits", "vitg"]
+            scored = []
+            for p in paths:
+                name = os.path.basename(p).lower()
+                score = 100
+                for i, tag in enumerate(order):
+                    if tag in name:
+                        score = i
+                        break
+                scored.append((score, p))
+            scored.sort(key=lambda x: x[0])
+            return [p for _, p in scored]
+
+        def _resolve_path(mp: str) -> str:
+            if isinstance(mp, str) and mp.strip().lower() == "auto":
+                mp = ""
+            if mp and os.path.isfile(mp):
+                return mp
+            search_dirs = []
+            if mp and os.path.isdir(mp):
+                search_dirs.append(mp)
+            base_dir = os.path.join(os.path.dirname(__file__), '..', 'depth-anything')
+            search_dirs.append(base_dir)
+            cand = []
+            for d in search_dirs:
+                try:
+                    for fn in os.listdir(d):
+                        if fn.lower().endswith('.pth') and 'depth_anything_v2' in fn.lower():
+                            cand.append(os.path.join(d, fn))
+                except Exception:
+                    pass
+            if cand:
+                return _prefer_order(cand)[0]
+            return mp
+
+        model_path = _resolve_path(model_path)
+    except Exception:
+        pass
     # Prefer our vendored implementation first
     try:
         from ...vendor.depth_anything_v2.dpt import DepthAnythingV2  # type: ignore
@@ -127,6 +168,22 @@ def _build_depth_map(image_bhwc: torch.Tensor, res: int, model_path: str, hires_
             d = torch.from_numpy(d)[None, None]  # 1,1,h,w
             d = F.interpolate(d, size=(H, W), mode='bilinear', align_corners=False)
             d = d[0, 0].to(device=dev, dtype=dtype)
+            # Heuristic de-banding: sometimes upstream returns column-wise banding
+            # Detect when column variance >> row variance, indicating vertical stripes
+            try:
+                with torch.no_grad():
+                    dr = d.clamp(0,1)
+                    # compute per-axis variance summaries in a small, cheap way
+                    vcol = torch.var(dr, dim=0).mean()
+                    vrow = torch.var(dr, dim=1).mean()
+                    if torch.isfinite(vcol) and torch.isfinite(vrow) and (vcol > 8.0 * (vrow + 1e-6)):
+                        # Apply mild horizontal smoothing only (reduce vertical banding)
+                        k = max(3, int(round(min(W, 21) // 2 * 2 + 1)))  # odd kernel up to ~21
+                        dr2 = F.avg_pool2d(dr.unsqueeze(0).unsqueeze(0), kernel_size=(1, k), stride=1, padding=(0, k//2))[0,0]
+                        # preserve global contrast by gentle blend
+                        d = (0.6 * dr + 0.4 * dr2).clamp(0,1)
+            except Exception:
+                pass
             d = d.clamp(0, 1)
             return d
         except Exception:

mod/mg_combinode.py

@@ -294,7 +294,7 @@ class MagicNodesCombiNode:
 
         # single clear at start is enough; avoid double-clearing here
 
-        # Применение цепочки LoRA
+        # Apply LoRA chain
         loras = [
             (use_lora_1, lora_1, strength_model_1, strength_clip_1),
             (use_lora_2, lora_2, strength_model_2, strength_clip_2),
@@ -345,7 +345,7 @@ class MagicNodesCombiNode:
         except Exception:
             pass
 
-        # Embeddings Positive и Negative
+        # Embeddings: Positive and Negative
         # Standard pipeline: optionally use a shared CLIP after clip_layer + CLIP-LoRA
         # Select CLIP source for encoding: pristine when standard pipeline is enabled
         src_clip = clip_clean if bool(standard_pipeline) else clip

pressets/mg_cade25.cfg

@@ -118,8 +118,8 @@ aq_attn: true
 seed: 0
 control_after_generate: randomize
 steps: 30
-cfg: 7
-denoise: 0.60
+cfg: 7.5
+denoise: 0.65
 sampler_name: ddim
 scheduler: MGHybrid
 iterations: 2
@@ -355,7 +355,7 @@ control_after_generate: randomize
 steps: 25
 cfg: 6.0
 #0.75
-denoise: 0.35
+denoise: 0.45
 sampler_name: ddim
 scheduler: MGHybrid
 iterations: 2

pressets/mg_controlfusion.cfg

@@ -12,7 +12,7 @@ pyra_low: 109
 pyra_high: 215
 pyra_resolution: 1024
 edge_thin_iter: 0
-edge_alpha: 0.75
+edge_alpha: 0.65
 edge_boost: 0.10
 smart_tune: false
 smart_boost: 0.62
@@ -20,8 +20,8 @@ smart_boost: 0.62
 # blend & strengths
 blend_mode: normal
 blend_factor: 0.35
-strength_pos: 0.47
-strength_neg: 0.91
+strength_pos: 0.45
+strength_neg: 1.10
 
 # schedule window
 start_percent: 0.000
@@ -38,10 +38,10 @@ stack_prev_control: false
 split_apply: false
 
 # split timings
-edge_start_percent: 0.000
-edge_end_percent: 0.600
-depth_start_percent: 0.000
-depth_end_percent: 1.000
+edge_start_percent: 0.200
+edge_end_percent: 0.400
+depth_start_percent: 0.200
+depth_end_percent: 0.400
 
 # multipliers & shape
 edge_strength_mul: 0.03
@@ -73,8 +73,8 @@ smart_boost: 0.84
 # blend & strengths
 blend_mode: normal
 blend_factor: 0.300
-strength_pos: 0.95
-strength_neg: 0.90
+strength_pos: 0.45
+strength_neg: 1.10
 
 # schedule window
 start_percent: 0.000
@@ -126,7 +126,7 @@ smart_boost: 0.56
 # blend & strengths
 blend_mode: normal
 blend_factor: 0.300
-strength_pos: 0.90
+strength_pos: 0.75
 strength_neg: 1.00
 
 # schedule window