flux-v2 v4 rev: fix attn_bias+ControlNet conditioning bug (multiply 0.35), bfl-lora to BFL spec (guidance 10, LoRA 0.85), calibration-validated

wip/flux-redux/flux-v2-redux-style-composition-bfl-lora.json

@@ -26,7 +26,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        "## FLUX Redux + official BFL Depth LoRA (v2 \u2014 A/B variant)\n\nv2 changes vs v1: portrait pre-pass (896x1152 fill/crop \u2014 critical here: with latent-concat\nconditioning any padding becomes image content), saved crop + depth map, steps 32, and\n**FluxGuidance 10 \u2192 4.0** (the v1 quality killer; BFL's guidance-10 rec is for their FULL depth\nmodel, not the LoRA \u2014 sweep showed lower is better). Structure looseness knob = LoRA strength\n(1.0 full grip, ~0.85 looser). No canny stacking possible in this method.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
+        "## FLUX Redux + official BFL Depth LoRA (v2 \u2014 A/B variant)\n\nv2 changes vs v1: portrait pre-pass (896x1152 fill/crop \u2014 critical here: with latent-concat\nconditioning any padding becomes image content), saved crop + depth map, steps 32, and\n**FluxGuidance 10.0 + LoRA 0.85** (BFL spec \u2014 the LoRA is distilled at guidance 10; earlier\nlower-guidance results were an artifact of the attn_bias bug). Redux multiply 0.35, never\nattn_bias (see style-composition notes). No canny stacking possible in this method.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
       ]
     },
     {
@@ -370,7 +370,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        4.0
+        10.0
       ]
     },
     {
@@ -461,8 +461,8 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        0.5,
-        "attn_bias"
+        0.35,
+        "multiply"
       ]
     },
     {
@@ -777,7 +777,7 @@
       "title": "BFL Depth LoRA",
       "widgets_values": [
         "flux1-depth-dev-lora.safetensors",
-        1.0
+        0.85
       ]
     },
     {

wip/flux-redux/flux-v2-redux-style-composition.json

@@ -26,7 +26,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        "## FLUX Redux + ControlNet \u2014 style + composition (v2)\n\nv2 changes vs v1: portrait pre-pass (any input \u2192 exact **896x1152** fill/crop, so depth map ==\nlatent, zero padding), saved input crops + depth maps for auditing, steps 32, and a bypassed\n**HIRES** group (1.5x latent upscale \u2192 second pass at denoise 0.30 with Redux conditioning\ncarried, ControlNet released \u2014 enable for ~1344x1728 finals; keeps aesthetic, adds detail).\n\nModels & knobs: unchanged from v1 (see `flux-redux-style-composition.json` notes). Redux\nattn_bias 0.5, Union-Pro-2.0 depth 0.7 / end 0.8, FluxGuidance **3.0** (sweep-validated; 2.5 softer/filmic, 3.5 punchier), euler/simple/32/CFG 1.\nBypassed groups: CANNY stack (0.35), ReduxAdvanced alt, turbo fast-preview, HIRES.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
+        "## FLUX Redux + ControlNet \u2014 style + composition (v2)\n\nv2 changes vs v1: portrait pre-pass (any input \u2192 exact **896x1152** fill/crop, so depth map ==\nlatent, zero padding), saved input crops + depth maps for auditing, steps 32, and a bypassed\n**HIRES** group (1.5x latent upscale \u2192 second pass at denoise 0.30 with Redux conditioning\ncarried, ControlNet released \u2014 enable for ~1344x1728 finals; keeps aesthetic, adds detail).\n\nModels & knobs: unchanged from v1 (see `flux-redux-style-composition.json` notes). Redux **multiply 0.35** \u2014 NEVER attn_bias with a ControlNet: its attention mask never reaches\nthe CN branch (code-verified), so base and CN fight and bodies deform. Union-Pro-2.0 depth\n0.7 / end 0.8 (drop end to 0.6 for more freedom), FluxGuidance **3.0** (sweep-validated; 2.5 softer/filmic, 3.5 punchier), euler/simple/32/CFG 1.\nBypassed groups: CANNY stack (0.35), ReduxAdvanced alt, turbo fast-preview, HIRES.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
       ]
     },
     {
@@ -469,8 +469,8 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        0.5,
-        "attn_bias"
+        0.35,
+        "multiply"
       ]
     },
     {

workflows/flux-v2-redux-style-composition-bfl-lora.json

@@ -26,7 +26,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        "## FLUX Redux + official BFL Depth LoRA (v2 \u2014 A/B variant)\n\nv2 changes vs v1: portrait pre-pass (896x1152 fill/crop \u2014 critical here: with latent-concat\nconditioning any padding becomes image content), saved crop + depth map, steps 32, and\n**FluxGuidance 10 \u2192 4.0** (the v1 quality killer; BFL's guidance-10 rec is for their FULL depth\nmodel, not the LoRA \u2014 sweep showed lower is better). Structure looseness knob = LoRA strength\n(1.0 full grip, ~0.85 looser). No canny stacking possible in this method.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
+        "## FLUX Redux + official BFL Depth LoRA (v2 \u2014 A/B variant)\n\nv2 changes vs v1: portrait pre-pass (896x1152 fill/crop \u2014 critical here: with latent-concat\nconditioning any padding becomes image content), saved crop + depth map, steps 32, and\n**FluxGuidance 10.0 + LoRA 0.85** (BFL spec \u2014 the LoRA is distilled at guidance 10; earlier\nlower-guidance results were an artifact of the attn_bias bug). Redux multiply 0.35, never\nattn_bias (see style-composition notes). No canny stacking possible in this method.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
       ]
     },
     {
@@ -370,7 +370,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        4.0
+        10.0
       ]
     },
     {
@@ -461,8 +461,8 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        0.5,
-        "attn_bias"
+        0.35,
+        "multiply"
       ]
     },
     {
@@ -777,7 +777,7 @@
       "title": "BFL Depth LoRA",
       "widgets_values": [
         "flux1-depth-dev-lora.safetensors",
-        1.0
+        0.85
       ]
     },
     {

workflows/flux-v2-redux-style-composition.json

@@ -26,7 +26,7 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        "## FLUX Redux + ControlNet \u2014 style + composition (v2)\n\nv2 changes vs v1: portrait pre-pass (any input \u2192 exact **896x1152** fill/crop, so depth map ==\nlatent, zero padding), saved input crops + depth maps for auditing, steps 32, and a bypassed\n**HIRES** group (1.5x latent upscale \u2192 second pass at denoise 0.30 with Redux conditioning\ncarried, ControlNet released \u2014 enable for ~1344x1728 finals; keeps aesthetic, adds detail).\n\nModels & knobs: unchanged from v1 (see `flux-redux-style-composition.json` notes). Redux\nattn_bias 0.5, Union-Pro-2.0 depth 0.7 / end 0.8, FluxGuidance **3.0** (sweep-validated; 2.5 softer/filmic, 3.5 punchier), euler/simple/32/CFG 1.\nBypassed groups: CANNY stack (0.35), ReduxAdvanced alt, turbo fast-preview, HIRES.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
+        "## FLUX Redux + ControlNet \u2014 style + composition (v2)\n\nv2 changes vs v1: portrait pre-pass (any input \u2192 exact **896x1152** fill/crop, so depth map ==\nlatent, zero padding), saved input crops + depth maps for auditing, steps 32, and a bypassed\n**HIRES** group (1.5x latent upscale \u2192 second pass at denoise 0.30 with Redux conditioning\ncarried, ControlNet released \u2014 enable for ~1344x1728 finals; keeps aesthetic, adds detail).\n\nModels & knobs: unchanged from v1 (see `flux-redux-style-composition.json` notes). Redux **multiply 0.35** \u2014 NEVER attn_bias with a ControlNet: its attention mask never reaches\nthe CN branch (code-verified), so base and CN fight and bodies deform. Union-Pro-2.0 depth\n0.7 / end 0.8 (drop end to 0.6 for more freedom), FluxGuidance **3.0** (sweep-validated; 2.5 softer/filmic, 3.5 punchier), euler/simple/32/CFG 1.\nBypassed groups: CANNY stack (0.35), ReduxAdvanced alt, turbo fast-preview, HIRES.\n\n**Optimal inputs** (both refs): 3:4 portrait, 1152x1536, sRGB 8-bit JPEG q90+/PNG, no\nwatermarks/text/borders, EXIF baked. Style ref: keep the style-defining content in the vertical\nmiddle (sigclip center-crops to a square \u2014 top/bottom ~12% are discarded). Composition ref:\nclear fg/bg depth separation. Any other input still works \u2014 the pre-pass normalizes to 896x1152\n(fill/crop) and the exact crop + depth map are saved next to the outputs for auditing."
       ]
     },
     {
@@ -469,8 +469,8 @@
         "ver": "0.27.0"
       },
       "widgets_values": [
-        0.5,
-        "attn_bias"
+        0.35,
+        "multiply"
       ]
     },
     {