Add files using upload-large-folder tool

loras/ZIT_D351_L1ght_V1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ef949a89411c824fd9bb8be0c0748943d07871d7f9743333c943af180cac06
+size 170128232

loras/ZIT_D351_L1ght_V1.safetensors.txt

@@ -0,0 +1,138 @@
+https://civitai.red/models/2349412/dezi-light-lora-for-indian-south-asian-faces-zit
+Details
+Type
+LoRA
+Stats
+566
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+
+381
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+
+774
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+
+Generation
+Not loaded
+
+Reviews
+Positive
+(28)
+
+Published
+Jan 30, 2026
+
+Base Model
+ZImageTurbo
+
+Training
+Steps: 1,500
+Epochs: 35
+Usage Tips
+Clip Skip: 2
+Strength: 0.6
+Hash
+Trigger Words
+d351 l1ght
+AIR
+civitai:
+2349412
+@
+2642504
++
+3065524
+This model was trained on Z-Image Turbo which otherwise creates 'standard' South Asian faces and skintones.
+
+Trigger word:
+
+d351 l1ght - lighter skintones
+
+Would love to see someone test this out!
+

loras/zit_stage1_masked_dpo_11100.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4395632e6ed627f7e89a102cf5a6cfbb0ce9f2605994de593ee50e301126d4f7
+size 42662256

loras/zit_stage1_masked_dpo_11100.safetensors.txt

@@ -0,0 +1,28 @@
+https://huggingface.co/F16/z-image-turbo-masked-dpo/tree/main?not-for-all-audiences=true
+
+
+Gated model
+
+You have been granted access to this model
+
+Masked Flow-DPO: Region-Aware Preference Optimization for Flow Matching Models
+
+Telegram
+Released Weights (Stage 1)
+
+The published LoKR adapter corrects the human anatomy of the Z-Image-Turbo base model. After applying this adapter:
+
+    Correct body proportions and shapes — limbs, torso, and extremities now have anatomically plausible geometry instead of the deformed outputs the base model produces
+    Improved fine details — localized anatomical structures and intimate details are rendered with correct morphology
+    No style or composition regression — corrections are spatially masked to the body region; background, lighting, and overall composition remain largely unchanged from the base model
+
+	
+Format 	LoKR (Kronecker low-rank adapter), .safetensors
+Base model 	Z-Image-Turbo (8-step distilled DiT)
+
+    Load this adapter on top of Z-Image-Turbo using any LoRA-compatible inference pipeline. No other changes to the base model or sampling config are needed.
+
+Masked Flow-DPO is a training framework that extends Flow-DPO (arXiv:2501.13918) with spatial masking, enabling preference optimization to focus on specific image regions rather than optimizing the entire frame globally. This is particularly effective when win/lose image pairs differ only in a localized area (e.g., anatomical correctness, facial details, object quality) while sharing identical backgrounds and composition.
+The Problem: Signal Dilution in Localized Corrections
+
+Standard DPO computes MSE loss over the entire latent space. When a win/lose pair differs only in a small region (~5-10% of pixels), the gradient signal from the meaningful difference is overwhelmed by noise from the vast identical area. The optimizer sees a weak, noisy preference signal and struggles to learn the intended correction.