Add NAG fallback and fp32 decode options, update presets

Introduces NAG fallback logic to both easy and hard modules for improved guidance when CrossAttention patch is inactive. Adds an option to force VAE decode output to fp32 in the easy module. Updates mg_cade25.cfg with new default values for steps, cfg, denoise, buffer, guidance, and other parameters to improve quality and stability.

mod/easy/mg_cade25_easy.py

@@ -1070,7 +1070,7 @@ def _scheduler_names():
         return ["normal", "MGHybrid"]
 
 
-def safe_decode(vae, lat, tile=512, ovlp=64):
+def safe_decode(vae, lat, tile=512, ovlp=128, to_fp32: bool = False):
     # Ensure we don't build autograd graphs during final decode steps
     with torch.inference_mode():
         h, w = lat["samples"].shape[-2:]
@@ -1091,6 +1091,12 @@ def safe_decode(vae, lat, tile=512, ovlp=64):
             out_cpu = out_cpu.to('cpu')
         except Exception:
             pass
+        # Optional: force fp32 decode output (after moving to CPU to save VRAM)
+        try:
+            if bool(to_fp32) and out_cpu.dtype != torch.float32:
+                out_cpu = out_cpu.float()
+        except Exception:
+            pass
         try:
             del out
         except Exception:
@@ -1324,7 +1330,8 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
                                 cfg_sched_type: str = "off", cfg_sched_min: float = 0.0, cfg_sched_max: float = 0.0,
                                 cfg_sched_gamma: float = 1.5, cfg_sched_u_pow: float = 1.0,
                                 cwn_enable: bool = True, alpha_c: float = 1.0, alpha_u: float = 1.0,
-                                agc_enable: bool = True, agc_tau: float = 2.8):
+                                agc_enable: bool = True, agc_tau: float = 2.8,
+                                nag_fb_enable: bool = False, nag_fb_scale: float = 4.0, nag_fb_tau: float = 2.5, nag_fb_alpha: float = 0.25):
 
     """Clone model and attach a cfg mixing function implementing RescaleCFG/FDG, CFGZero*/FD, or hybrid ZeResFDG.
     guidance_mode: 'default' | 'RescaleCFG' | 'RescaleFDG' | 'CFGZero*' | 'CFGZeroFD' | 'ZeResFDG'
@@ -1361,6 +1368,25 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
         sigma = args.get("sigma", None)
         x_orig = args.get("input", None)
 
+        # --- NAG fallback in noise-space (when CrossAttention patch is inactive) ---
+        if bool(nag_fb_enable):
+            try:
+                active = bool(getattr(sa_patch, "_nag_patch_active", False))
+            except Exception:
+                active = False
+            if not active:
+                try:
+                    phi = float(nag_fb_scale); tau = float(nag_fb_tau); a = float(nag_fb_alpha)
+                    g = cond * phi - uncond * (phi - 1.0)
+                    def _l1(x):
+                        return torch.sum(torch.abs(x), dim=(1,2,3), keepdim=True).clamp_min(1e-6)
+                    s_pos = _l1(cond); s_g = _l1(g)
+                    scale = (s_pos * tau) / s_g
+                    g = torch.where((s_g > s_pos * tau), g * scale, g)
+                    cond = g * a + cond * (1.0 - a)
+                except Exception:
+                    pass
+
         # Local spatial gain from CURRENT_ONNX_MASK_BCHW, resized to cond spatial size
         def _local_gain_for(hw):
             if not bool(use_local_mask):
@@ -2242,6 +2268,9 @@ class ComfyAdaptiveDetailEnhancer25:
         cfg_sched_gamma = float(pv("cfg_sched_gamma", 1.5))
         cfg_sched_u_pow = float(pv("cfg_sched_u_pow", 1.0))
 
+        # VAE decode: allow forcing fp32 output (default false)
+        vae_decode_fp32 = bool(pv("vae_decode_fp32", False))
+
         # CWN + AGC defaults (hidden in Easy; controlled via presets)
         cwn_enable = bool(pv("cwn_enable", True))
         alpha_c = float(pv("alpha_c", 1.0))
@@ -2277,7 +2306,7 @@ class ComfyAdaptiveDetailEnhancer25:
         except Exception:
             pass
 
-        image = safe_decode(vae, latent)
+        image = safe_decode(vae, latent, to_fp32=bool(vae_decode_fp32))
         # allow user cancel right after initial decode
         model_management.throw_exception_if_processing_interrupted()
 
@@ -2502,7 +2531,8 @@ class ComfyAdaptiveDetailEnhancer25:
                       cfg_sched_type=str(cfg_sched), cfg_sched_min=float(cfg_sched_min), cfg_sched_max=float(cfg_sched_max),
                       cfg_sched_gamma=float(cfg_sched_gamma), cfg_sched_u_pow=float(cfg_sched_u_pow),
                       cwn_enable=bool(cwn_enable), alpha_c=float(alpha_c), alpha_u=float(alpha_u),
-                      agc_enable=bool(agc_enable), agc_tau=float(agc_tau)
+                      agc_enable=bool(agc_enable), agc_tau=float(agc_tau),
+                      nag_fb_enable=bool(use_nag), nag_fb_scale=float(nag_scale), nag_fb_tau=float(nag_tau), nag_fb_alpha=float(nag_alpha)
                   )
                 # check once more right before the loop starts
                 model_management.throw_exception_if_processing_interrupted()
@@ -2549,7 +2579,7 @@ class ComfyAdaptiveDetailEnhancer25:
                         try:
                             import os
                             models_dir = os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), "models")
-                            img_preview = safe_decode(vae, current_latent)
+                            img_preview = safe_decode(vae, current_latent, to_fp32=bool(vae_decode_fp32))
                             # Set toggles for this iteration
                             globals()["_ONNX_DEBUG"] = bool(onnx_debug)
                             globals()["_ONNX_COUNT_DEBUG"] = True  # force counts ON for debugging session
@@ -2583,7 +2613,7 @@ class ComfyAdaptiveDetailEnhancer25:
                     # CF edge mask (from current image) and fusion (only when enabled)
                     if bool(seg_use_cf_edges):
                         try:
-                            img_prev2 = safe_decode(vae, current_latent)
+                            img_prev2 = safe_decode(vae, current_latent, to_fp32=bool(vae_decode_fp32))
                             em2 = _build_cf_edge_mask_from_step(img_prev2, str(preset_step))
                             if em2 is not None:
                                 if onnx_mask_last is None:
@@ -2680,7 +2710,7 @@ class ComfyAdaptiveDetailEnhancer25:
                     try:
                         do_local_refine = False  # disable local best-of-2 by default
                         if do_local_refine:
-                            img_roi = safe_decode(vae, current_latent)
+                            img_roi = safe_decode(vae, current_latent, to_fp32=bool(vae_decode_fp32))
                             roi = _clipseg_build_mask(img_roi, "hand | hands | face", preview=max(192, int(clipseg_preview//2)), threshold=0.40, blur=5.0, dilate=2, gain=1.0)
                             if roi is None and onnx_mask_last is not None:
                                 roi = torch.clamp(onnx_mask_last, 0.0, 1.0)
@@ -2710,13 +2740,13 @@ class ComfyAdaptiveDetailEnhancer25:
                                 lat_a, = nodes.common_ksampler(
                                     sampler_model, s1, micro_steps, current_cfg, sampler_name, scheduler,
                                     positive, negative, lat_in_a, denoise=micro_denoise)
-                                img_a = safe_decode(vae, lat_a)
+                                img_a = safe_decode(vae, lat_a, to_fp32=bool(vae_decode_fp32))
                                 # Candidate B
                                 lat_in_b = {"samples": current_latent["samples"].clone()}
                                 lat_b, = nodes.common_ksampler(
                                     sampler_model, s2, micro_steps, current_cfg, sampler_name, scheduler,
                                     positive, negative, lat_in_b, denoise=micro_denoise)
-                                img_b = safe_decode(vae, lat_b)
+                                img_b = safe_decode(vae, lat_b, to_fp32=bool(vae_decode_fp32))
 
                                 # Score inside ROI
                                 def _roi_stats(img, roi_mask):
@@ -2962,7 +2992,7 @@ class ComfyAdaptiveDetailEnhancer25:
                     except Exception:
                         pass
 
-                    image = safe_decode(vae, current_latent)
+                    image = safe_decode(vae, current_latent, to_fp32=bool(vae_decode_fp32))
                     # and again after decode before post-processing
                     model_management.throw_exception_if_processing_interrupted()
 
@@ -3459,7 +3489,7 @@ def _smart_seed_select(model,
                     model, int(sd), int(probe_steps), float(cfg), str(sampler_name), str(scheduler),
                     positive, negative, lat_in, denoise=float(min(denoise, 0.65))
                 )
-                img = safe_decode(vae, lat_out)
+                img = safe_decode(vae, lat_out, to_fp32=bool(vae_decode_fp32))
                 # and again right after decode
                 model_management.throw_exception_if_processing_interrupted()
                 # Base score: edge density toward a target + low speckle + balanced exposure
@@ -3598,3 +3628,4 @@ def _interruptible_ksampler(model, seed, steps, cfg, sampler_name, scheduler,
     out = {**latent}
     out["samples"] = samples
     return (out,)
+

mod/hard/mg_cade25.py

@@ -971,7 +971,9 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
                                 eps_scale_enable: bool = False, eps_scale: float = 0.0,
                                 # NEW: CWN + AGC for Hard node too
                                 cwn_enable: bool = True, alpha_c: float = 1.0, alpha_u: float = 1.0,
-                                agc_enable: bool = True, agc_tau: float = 2.8):
+                                agc_enable: bool = True, agc_tau: float = 2.8,
+                                # NAG fallback
+                                nag_fb_enable: bool = False, nag_fb_scale: float = 4.0, nag_fb_tau: float = 2.5, nag_fb_alpha: float = 0.25):
 
     """Clone model and attach a cfg mixing function implementing RescaleCFG/FDG, CFGZero*/FD, or hybrid ZeResFDG.
     guidance_mode: 'default' | 'RescaleCFG' | 'RescaleFDG' | 'CFGZero*' | 'CFGZeroFD' | 'ZeResFDG'
@@ -1008,6 +1010,26 @@ def _wrap_model_with_guidance(model, guidance_mode: str, rescale_multiplier: flo
         sigma = args.get("sigma", None)
         x_orig = args.get("input", None)
 
+        # NAG fallback (noise-space) when CrossAttention patch inactive
+        if bool(nag_fb_enable):
+            try:
+                from . import mg_sagpu_attention as _sa
+                active = bool(getattr(_sa, "_nag_patch_active", False))
+            except Exception:
+                active = False
+            if not active:
+                try:
+                    phi = float(nag_fb_scale); tau = float(nag_fb_tau); a = float(nag_fb_alpha)
+                    g = cond * phi - uncond * (phi - 1.0)
+                    def _l1(x):
+                        return torch.sum(torch.abs(x), dim=(1,2,3), keepdim=True).clamp_min(1e-6)
+                    s_pos = _l1(cond); s_g = _l1(g)
+                    scale = (s_pos * tau) / s_g
+                    g = torch.where((s_g > s_pos * tau), g * scale, g)
+                    cond = g * a + cond * (1.0 - a)
+                except Exception:
+                    pass
+
         # Local spatial gain from CURRENT_ONNX_MASK_BCHW, resized to cond spatial size
         def _local_gain_for(hw):
             if not bool(use_local_mask):
@@ -1762,7 +1784,8 @@ class ComfyAdaptiveDetailEnhancer25:
                       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),
                       cwn_enable=bool(cwn_enable), alpha_c=float(alpha_c), alpha_u=float(alpha_u),
-                      agc_enable=bool(agc_enable), agc_tau=float(agc_tau)
+                      agc_enable=bool(agc_enable), agc_tau=float(agc_tau),
+                      nag_fb_enable=bool(use_nag), nag_fb_scale=float(nag_scale), nag_fb_tau=float(nag_tau), nag_fb_alpha=float(nag_alpha)
                   )
                 # early interruption check before starting the loop
                 try:

pressets/mg_cade25.cfg

@@ -1,4 +1,5 @@
-# MagicNodes CADE 2.5 presets
+
+# MagicNodes CADE 2.5 presets
 
 [step1]
 # core
@@ -280,14 +281,15 @@ aq_alpha: 2.0
 
 aq_attn: true
 
+vae_decode_fp32: true
 
 [step3]
 # core
 seed: 0
 control_after_generate: randomize
-steps: 16
-cfg: 7.0
-denoise: 0.55
+steps: 20
+cfg: 6.0
+denoise: 0.46
 sampler_name: ddim
 scheduler: MGHybrid
 iterations: 2
@@ -307,13 +309,13 @@ latent_compare: true
 
 # latent buffer (internal)
 latent_buffer: true
-lb_inject: 0.25
+lb_inject: 0.10
 lb_ema: 0.75
 lb_every: 1
 lb_anchor_every: 4
 lb_masked: true
-lb_rebase_thresh: 0.10
-lb_rebase_rate: 0.25
+lb_rebase_thresh: 0.18
+lb_rebase_rate: 0.10
 
 # detail controls
 #ids_strength: 0.25
@@ -321,7 +323,7 @@ ids_strength: 0.55
 upscale_method: lanczos
 scale_by: 1.20
 scale_delta: 0.10
-noise_offset: 0.10
+noise_offset: 0.03
 threshold: 1.000
 #0.135
 Sharpnes_strenght: 0.185
@@ -330,35 +332,35 @@ accumulation: fp32+fp32
 # reference clean
 reference_clean: true
 ref_preview: 512
-ref_threshold: 0.020
+ref_threshold: 0.320
 ref_cooldown: 2
 
 # cfg schedule (internal)
 #cfg_sched: off | cosine | warmup | u
 cfg_sched: warmup
-cfg_sched_min: 5.8
-cfg_sched_max: 7.8
+cfg_sched_min: 5.0
+cfg_sched_max: 6.5
 cfg_sched_gamma: 1.5
 cfg_sched_u_pow: 1.2
 
 
 # guidance
 guidance_mode: ZeResFDG
-rescale_multiplier: 0.75
-momentum_beta: 0.15
-cfg_curve: 0.85
-perp_damp: 0.80
+rescale_multiplier: 0.95
+momentum_beta: 0.65
+cfg_curve: 1.15
+perp_damp: 0.90
 
 # CWN/AGC
 cwn_enable: true
-alpha_c: 1.0
-alpha_u: 1.0
+alpha_c: 1.09
+alpha_u: 0.8
 agc_enable: true
-agc_tau: 2.8
+agc_tau: 1.2
 
 # NAG
 use_nag: true
-nag_scale: 4.0
+nag_scale: 3.6
 nag_tau: 2.50
 nag_alpha: 0.25
 
@@ -367,16 +369,16 @@ use_zero_init: false
 zero_init_steps: 0
 
 # FDG / ZE thresholds
-fdg_low: 0.15
-fdg_high: 1.15
-fdg_sigma: 1.10
+fdg_low: 0.45
+fdg_high: 1.45
+fdg_sigma: 1.80
 ze_res_zero_steps: 6
 ze_adaptive: true
-ze_r_switch_hi: 0.85
+ze_r_switch_hi: 0.55
 ze_r_switch_lo: 0.25
 fdg_low_adaptive: true
-fdg_low_min: 0.45
-fdg_low_max: 0.85
+fdg_low_min: 0.15
+fdg_low_max: 1.85
 fdg_ema_beta: 0.45
 
 
@@ -386,7 +388,7 @@ muse_blend_strength: 0.24
 
 # eps scale
 eps_scale_enable: true
-eps_scale: 0.0025
+eps_scale: 0.0225
 
 # CLIPSeg
 clipseg_enable: true
@@ -401,7 +403,7 @@ clipseg_ref_gate: true
 clipseg_ref_threshold: 0.005
 
 # polish
-polish_enable: false
+polish_enable: true
 polish_keep_low: 0.40
 polish_edge_lock: 0.20
 polish_sigma: 1.0
@@ -411,8 +413,8 @@ polish_keep_low_ramp: 0.20
 # mid-frequency stabilizer (hands/objects scale)
 midfreq_enable: true
 #midfreq_gain: 0.15
-midfreq_gain: 0.65
-midfreq_sigma_lo: 0.90
+midfreq_gain: 0.45
+midfreq_sigma_lo: 0.80
 midfreq_sigma_hi: 2.10
 
 # QSilk-AQClip-Lite (adaptive latent clipping)
@@ -423,14 +425,16 @@ aq_alpha: 2.0
 
 aq_attn: true
 
+vae_decode_fp32: true
+
 [step4]
 # core
 seed: 0
 control_after_generate: randomize
-steps: 16
+steps: 20
 cfg: 6
 #0.75
-denoise: 0.39
+denoise: 0.46
 sampler_name: ddim
 scheduler: MGHybrid
 iterations: 2
@@ -453,7 +457,7 @@ latent_buffer: true
 lb_inject: 0.30
 lb_ema: 0.75
 lb_every: 1
-lb_anchor_every: 6
+lb_anchor_every: 3
 lb_masked: true
 lb_rebase_thresh: 0.10
 lb_rebase_rate: 0.25
@@ -464,7 +468,7 @@ upscale_method: lanczos
 # 1.55
 scale_by: 1.50
 scale_delta: 0.1
-noise_offset: 0.0035
+noise_offset: 0.0135
 threshold: 1.000
 Sharpnes_strenght: 0.24
 accumulation: fp32+fp32
@@ -478,7 +482,7 @@ ref_cooldown: 2
 # cfg schedule (internal)
 #cfg_sched: off | cosine | warmup | u
 cfg_sched: cosine
-cfg_sched_min: 3.9
+cfg_sched_min: 4.9
 cfg_sched_max: 8.0
 cfg_sched_gamma: 1.5
 cfg_sched_u_pow: 1.2
@@ -493,10 +497,10 @@ perp_damp: 0.85
 
 # CWN/AGC
 cwn_enable: true
-alpha_c: 1.0
+alpha_c: 1.03
 alpha_u: 1.0
 agc_enable: true
-agc_tau: 2.2
+agc_tau: 3.2
 
 # NAG
 use_nag: true
@@ -510,7 +514,7 @@ zero_init_steps: 0
 
 # FDG / ZE thresholds
 fdg_low: 0.35
-fdg_high: 0.90
+fdg_high: 1.10
 fdg_sigma: 1.20
 ze_res_zero_steps: 10
 ze_adaptive: true
@@ -544,7 +548,7 @@ clipseg_ref_threshold: 0.005
 seg_use_cf_edges: true
 
 # polish
-polish_enable: false
+polish_enable: true
 polish_keep_low: 0.45
 polish_edge_lock: 0.12
 polish_sigma: 0.8
@@ -572,5 +576,4 @@ aq_attn: true
 # KV pruning (self-attention speedup)
 kv_prune_enable: false
 kv_keep: 0.95
-kv_min_tokens: 256
-
+kv_min_tokens: 256