May 7 13:27 add Photoshope ClownsharK workflow Apr 24 23:35 Snapshot all workflows from live instance before shutdown (48 files) 23:10 Oz-zImage: restore active latents pool in Oz Adv Loader (was empty -> sampler collapse) 22:57 sanitize: replace stale ComfyUI_temp_* in LoadImage with permanent filename 22:57 sanitize: replace stale ComfyUI_temp_* in LoadImage with permanent filename 21:33 Kiara Z MAX: bypass Stage1 ControlNet, lower Stage2 strength 1.0->0.5 (dual strength=1 was collapsing output to dark) 21:25 Kiara Z MAX: enable ModelPatchLoader (150) and Set_Controlnet Model (480) so Stage1/Stage2 ControlNet receive weights 21:23 Kiara Z MAX: wire ControlNet cascade (LoRA -> Stage1 -> Stage2 -> Set_Model), fix Stage1 unwired model, fix orphaned SetNodes 19:54 ShotSplit: bypass ImageCompositeMasked 211 (keep full SR face/hair, not original-paste) 16:10 ShotSplit: replace Wonder2 node with SeedVR2 loader + wonder2 finetune DiT (same output, proper 2.5x upscale) 14:58 SeedVR2 standalone: fix input/output names (image/dit/vae, SEEDVR2_DIT, SEEDVR2_VAE) 14:34 Add standalone SeedVR2 upscale workflow (LoadImage -> SeedVR -> Save) 14:28 Wonder2: scale 4x->2x, cfg 1->2.0 14:28 Wonder2: scale 4x->2x, cfg 1->2.0 (4x bicubic start too blurry for DiT to recover) 14:20 workflow_wonder2: steps 1->15 to fix soft 4x upscale 14:18 ShotSplit V2.5: Wonder2 steps 1->15 to fix soft/blurry 4x upscale 14:12 ShotSplit V2.5: hair grow mask 30->200, USE MASK INPAINT default False (pure T2I path) 11:00 V2.5: bypass stale ozW_SynthesizeAuthenticMetadata node 402 10:39 V2.5: wire ShotSplitter.clip_paths -> Load Frames (sampled), fix VRAM_Debug slot Apr 22 10:40 add Z-Image Turbo darkHUB Photoshope workflow Apr 21 20:32 Sync all workflows from vast instance (42 files) 17:34 Add V2.5: mask/T2I toggle + pose toggle + Wonder2 + batch-of-5 Grok picker + VRAM unload nodes 07:36 V2.3 APPROACH A: mask-inpaint (face+hair union) preserves source composition exactly while replacing head with full papeech LoRA (black curly bob + papeech face). No ControlNet needed. 20 iterations to find winning combo. 00:52 V2.3 FINAL: denoise 0.75, LoRA m=1.0 c=1.0 (full papeech: black curly bob, original face), canny strength 0.35 — preserves pose/composition from source but all style from LoRA 00:32 V2.3 FINAL: LoRA strength_clip 1.0 -> 0.5 — preserves source hairstyle/color while keeping papeech facial identity (long wavy red hair vs curly bob) 00:21 V2.3 WIN: img2img (VAEEncode source frame) + ZCN Canny @ strength 0.3 + denoise 0.45 preserves source composition (pose/angle/framing). Iter-tested via API loop. Apr 20 23:45 V2.3: swap OpenPose -> DWPose (better skeleton detection on cropped selfies, yolox_l bbox + dw-ll pose estimator) 23:39 V2.3: fix OpenposePreprocessor widgets order (detect_hand, detect_body, detect_face, resolution=1024, scale_stick_for_xinsr_cn=disable) 23:34 V2.3: switch ZImageFunControlnet image from depth -> OpenPose (crisp skeleton > blurry depth for pose control); keep depth as comparison preview; bump depth resolution 518->1024 23:29 V2.3: move Depth/Frame previews near sampler, add SaveImage for depth mask 23:23 V2.3: batch 5->2 (OOM with ControlNet patch), add Depth+BestFrame previews, bump strength 0.65->0.85 23:05 V2.3 fix #2: ZImageFunControlnet now uses flux_ae.safetensors VAE (16ch) instead of SDXL VAE (4ch) — fixes matmul shape mismatch 23:00 V2.3 fix: use ZImageFunControlnet (model patch) instead of stock ControlNetLoader; model moved to model_patches/ 22:54 V2.3: add Z-Image ControlNet branch (BestFramePicker -> DepthAnythingV2 -> Fun-Controlnet-Union @ strength 0.65) for pose/angle grounding 20:36 ShotSplitter thresholds: transnet 0.5->0.3, pyscenedetect 3.0->2.0 (detected real hard cut at frame 173 missed by defaults) 19:55 Model split: reasoning for ShotAnalyzer + wan_pos/neg_zh; fast for detailers/upscaler/picker 19:50 Default Grok model -> grok-4.20-beta-0309-reasoning everywhere 15:01 V2.2: auto-wired GrokPromptWriter images+character_reference inputs 14:58 V2.2: Grok-driven detailer/upscaler/Wan 中文 prompts via ozW_GrokPromptWriter 13:45 V2.1: reconnect dead branches — #04 Upscale/NSFW now feeds #06 Hands (was reading from #02 Face, skipping upscale) 13:22 V2.1: Wan ref_images now uses fully post-processed output (#16 Multi-Compression) matching #17 final save 13:07 Add V2.1: batch_size=5 + Grok best-image picker (cumulative score) Apr 14 21:05 8-agent fix: match zImage 2.0 ref chain. Bypass USD pass (210-218), bypass ColorMatch branch, activate Lips Detailer, restore 543-537-548 chain order, FaceDetailer 644 guide_size 1920, Wan ref from clean 498 (not dirtied 551) 20:40 Reorder ref chain: FaceDetailer -> GLCM (native res) -> USD (1920) -> dirtying -> Wan. Fixes GLCM second-order stat corruption on upscaler output 20:14 Match zImage 2.0 ref chain: unbypass GLCM_Normalize, revert detailer to 1920/12steps/res_2s/0.42 (GLCM was anchor not breaker) 19:59 Fix cartoonish eyes: detailer guide_size 1920->512, steps 12->20, denoise 0.42->0.25, euler/karras, natural prompt 19:14 Rewire StitchReel.trigger from WanVideoDecode.images to TSVideoCombine(2977).Filenames 18:54 Bypass GLCM_Normalize — breaks realism, straight to RealisticNoise grain 17:43 Reroute UltimateSDUpscale -> GLCM_Normalize (feeds Wan.ref_images with 1920px HQ reference) 17:21 Perf: lower source res 480x832, frame_window=81, bypass redundant TSVideoCombine 2972 
