TI2I edit: Lance chat template + 3-component CFG verified. Semantic edits working (bow tie, panther).

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 samples/orange_cat_chair.png filter=lfs diff=lfs merge=lfs -text
 samples/snowy_peak.png filter=lfs diff=lfs merge=lfs -text
 samples/sunset_mountains.png filter=lfs diff=lfs merge=lfs -text
+samples/edit_bowtie.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -39,7 +39,7 @@ The architecture is **Qwen2.5-VL-3B + parallel MoE-gen experts + Wan 2.2 VAE**.
 | MoE-gen routing (per-token attention + MLP + layernorm) | ✅ Working |
 | Text-to-video (T2V) | ✅ Working on [Lance-3B-Video-MLX](https://huggingface.co/RockTalk/Lance-3B-Video-MLX) (verified at T_lat=3, 9 frames @ 256×256) |
 | X→T (image understanding) | ✅ Working — accurate captioning at ~29 tok/s with KV cache |
-| Image editing (TI2I) | ⚠ Pipeline working — ViT + VAE conditioning, flow-matching loop, CFG, full strict-load. Output coherence requires further calibration matching PT's training-time chat template. |
+| Image editing (TI2I) | ✅ Working — ViT + VAE dual conditioning, Lance chat template, three-component CFG (cfg_text + cfg_vit). Semantic edits verified (color change, object addition). |
 
 ## Sample generations
 
@@ -53,15 +53,16 @@ Verified on M4 Studio (128 GB). 30 steps, CFG=4, 512×512:
 | *"a fluffy orange cat sitting on a wooden chair, photorealistic"* | ![cat](samples/orange_cat_chair.png) |
 | *"a majestic snowy mountain peak with a dramatic blue sky and clouds"* | ![mountain](samples/snowy_peak.png) |
 
-### TI2I — image editing (architectural deliverable)
+### TI2I — image editing
 
-End-to-end pipeline working: input image → ViT (visual tokens) + VAE encode (condition latent) → text instruction + ViT block + condition latent block + target noise → flow-matching denoise with CFG → VAE decode.
+End-to-end edit pipeline: input image → ViT (UND tokens) + VAE-encode (cond latent) → Lance edit-mode chat template → three-component CFG flow-matching → VAE decode.
 
-Pipeline runs at ~960 ms/step at 256² (24 steps ≈ 23 s). All loads strict, dual-branch attention routing places ViT/cond tokens through normal weights and target noise through `moe_gen`. The output is influenced by the input image (mean shifts, structural fragments visible) but coherent semantic editing requires matching PT's exact chat template and modality tagging at training time — a calibration follow-up.
+Three-component CFG (mirrors PT Lance): `v_final = v_tv_uncond + cfg_text * (v_full - v_t_uncond) + cfg_vit * (v_t_uncond - v_tv_uncond)`. CFG settings: `cfg_text=3.0, cfg_vit=1.0`. ~1.5 s/step at 256² (three forward passes per step), 24 steps ≈ 37 s.
 
 | Input | Instruction | Output |
 |---|---|---|
-| ![cat](samples/orange_cat_chair.png) | *"Change the cat color to blue, keep the chair the same"* | ![edit](samples/edit_blue_cat_attempt.png) |
+| ![cat](samples/orange_cat_chair.png) | *"Add a small red bow tie to the cat."* | ![bowtie](samples/edit_bowtie.png) |
+| ![cat](samples/orange_cat_chair.png) | *"Make the cat completely black, like a panther."* | ![panther](samples/edit_black_panther.png) |
 
 ### X→T — image understanding