#220 manifest: add on-device IP-encode models to v75/v81 sets + ip_encode_models block

manifest.json

@@ -14,7 +14,8 @@
     "source=ours: clean-room RealVisXL -> W8A16 -> QAIRT 2.40 conversion, drop-in for our libsdxl_qnn_driver (graph name 'model', host combined-emb input).",
     "verified=true = run on real silicon and confirmed coherent (no body-horror). V75 (Lenovo .69) + V81 (RedMagic .82) device-verified. v73/v79 offline-validated, awaiting handsets.",
     "The 4 per-arch bins (unet/text_encoder_1_fp16/text_encoder_2_fp16/sdxl_emb_mlp) are the only RealVisXL swap; taesdxl_decoder.mnn (universal SDXL tiny-VAE) + the dual CLIP tokenizers + config are checkpoint-agnostic and byte-identical to the prior set.",
-    "#208 reference-image (IP-Adapter): per-arch IP-Adapter context bins (unet_ip.bin) added under ip/v75 + ip/v81 (each set's files map). Identity-agnostic (k_ip/v_ip computed per-render on CPU). V75 device-verified (ArcFace 0.417); V81 compiled, on-device-validation pending. Base sets + driver_required_files UNCHANGED."
+    "#208 reference-image (IP-Adapter): per-arch IP-Adapter context bins (unet_ip.bin) added under ip/v75 + ip/v81 (each set's files map). Identity-agnostic (k_ip/v_ip computed per-render on CPU). V75 device-verified (ArcFace 0.417); V81 compiled, on-device-validation pending. Base sets + driver_required_files UNCHANGED.",
+    "#220 NPU IP-encode productionized: the THREE on-device CPU IP-encode models (clip_vision_vith.mnn + its .weight sidecar, ip_resampler.mnn, ip_kv_proj.bin) published under ip/ and added to the v75+v81 sets' files maps (OPTIONAL set; absent -> plain render). They feed the rebuilt libsdxl_qnn_driver.so --encode_ref mode (#215, recompiled WITH ip_ref_encoder.cpp) so reference-image identity works fully from the shipped APK + the on-device-downloaded bundle. Base sets, unet_ip bins, and driver_required_files UNCHANGED (additive only)."
   ],
   "arch_to_socs": {
     "73": {
@@ -169,6 +170,26 @@
           "size": 2643410944,
           "md5": "ed58c09c834409d2fafbef8fb61fbad4"
         },
+        "clip_vision_vith.mnn": {
+          "path": "ip/clip_vision_vith.mnn",
+          "size": 413672,
+          "md5": "21b2b137a9ce2e2744a95310ca4ebc43"
+        },
+        "clip_vision_vith.mnn.weight": {
+          "path": "ip/clip_vision_vith.mnn.weight",
+          "size": 2444349440,
+          "md5": "6ff037e79376fee820c11505f796fd37"
+        },
+        "ip_resampler.mnn": {
+          "path": "ip/ip_resampler.mnn",
+          "size": 325673960,
+          "md5": "1dd445e5477c321ebd89faf691b8cfd6"
+        },
+        "ip_kv_proj.bin": {
+          "path": "ip/ip_kv_proj.bin",
+          "size": 1363149100,
+          "md5": "36fa9e46b8c7de6712546426245a6d44"
+        },
         "text_encoder_1_fp16.bin": {
           "path": "v75/text_encoder_1_fp16.bin",
           "size": 234770432,
@@ -340,6 +361,26 @@
           "size": 2648719360,
           "md5": "80a8f1342ed20aed802eb3eee45219a3"
         },
+        "clip_vision_vith.mnn": {
+          "path": "ip/clip_vision_vith.mnn",
+          "size": 413672,
+          "md5": "21b2b137a9ce2e2744a95310ca4ebc43"
+        },
+        "clip_vision_vith.mnn.weight": {
+          "path": "ip/clip_vision_vith.mnn.weight",
+          "size": 2444349440,
+          "md5": "6ff037e79376fee820c11505f796fd37"
+        },
+        "ip_resampler.mnn": {
+          "path": "ip/ip_resampler.mnn",
+          "size": 325673960,
+          "md5": "1dd445e5477c321ebd89faf691b8cfd6"
+        },
+        "ip_kv_proj.bin": {
+          "path": "ip/ip_kv_proj.bin",
+          "size": 1363149100,
+          "md5": "36fa9e46b8c7de6712546426245a6d44"
+        },
         "text_encoder_1_fp16.bin": {
           "path": "v81/text_encoder_1_fp16.bin",
           "size": 234754048,
@@ -465,10 +506,39 @@
         "path": "ip/ip_kv_proj_weights.npz",
         "size": 1363183986,
         "md5": "5ae6610c7f1adf62ff58f7b7751bf508",
-        "note": "The 70x to_k_ip/to_v_ip Linear weights ([inner,2048]) for the CPU IP side: keys k_0..k_69 / v_0..v_69, plus inner[70] + heads[70]. The on-device encode (CLIP-ViT-H + plus-face Resampler) consumes these to produce the per-render k_ip/v_ip blob. NB the CLIP-ViT-H encoder + plus-face adapter weights are the standard public h94/IP-Adapter SDXL plus-face files (ip_adapter_clip_vision_vith + ip-adapter-plus-face_sdxl_vit-h) and are not re-hosted here.",
+        "note": "The 70x to_k_ip/to_v_ip Linear weights ([inner,2048]) for the CPU IP side (.npz provenance/source form): keys k_0..k_69 / v_0..v_69, plus inner[70] + heads[70]. The on-device encode (CLIP-ViT-H + plus-face Resampler) consumes the packed ip_kv_proj.bin form (see ip_encode_models) to produce the per-render k_ip/v_ip blob.",
         "inner_dims": "0..3=640, 4..53=1280, 54..59=640, 60..69=1280 (graph order; matches ImageGenerationSdxlNpuIpKv.INNER_DIMS)"
       }
     },
+    "ip_encode_models": {
+      "note": "#220 — the THREE on-device CPU IP-encode MNN/bin models (+ CLIP-vision .weight sidecar) that turn the attached reference image into the per-render k_ip/v_ip blob, ALL on CPU via MNN inside the driver's --encode_ref mode (no DSP, identity-agnostic graphs). Arch-agnostic (pure CPU), so the same files serve every set; each per-arch set's `files` map references them under `ip/`. The app's ImageGenerationSdxlNpuDownloadPlan treats them as OPTIONAL bundle files (downloaded with the IP add-on, never required; absent -> reference silently degrades to a plain render). V75 device-verified end-to-end (ArcFace ~0.44 from the shipped libsdxl_qnn_driver.so --encode_ref + these on-device-downloaded models).",
+      "files": {
+        "clip_vision_vith.mnn": {
+          "path": "ip/clip_vision_vith.mnn",
+          "size": 413672,
+          "md5": "21b2b137a9ce2e2744a95310ca4ebc43",
+          "note": "CLIP-ViT-H vision graph (penultimate-hidden output [1,257,1280]); MNN large-model form -> carries the clip_vision_vith.mnn.weight external-weights sidecar which MUST sit next to it."
+        },
+        "clip_vision_vith.mnn.weight": {
+          "path": "ip/clip_vision_vith.mnn.weight",
+          "size": 2444349440,
+          "md5": "6ff037e79376fee820c11505f796fd37",
+          "note": "MNN external-weights sidecar for clip_vision_vith.mnn (must co-locate)."
+        },
+        "ip_resampler.mnn": {
+          "path": "ip/ip_resampler.mnn",
+          "size": 325673960,
+          "md5": "1dd445e5477c321ebd89faf691b8cfd6",
+          "note": "Plus-face Resampler (IPAdapterPlusImageProjection) -> 16 image-prompt tokens [1,16,2048]."
+        },
+        "ip_kv_proj.bin": {
+          "path": "ip/ip_kv_proj.bin",
+          "size": 1363149100,
+          "md5": "36fa9e46b8c7de6712546426245a6d44",
+          "note": "Packed 70x to_k_ip/to_v_ip Linear weights (driver-consumable form of ip_kv_proj_weights.npz, magic 'IPKW'); the encoder applies these to the resampler tokens to produce the IPKV blob."
+        }
+      }
+    },
     "blob_format": "IPKV LE: magic 0x564B5049 'IPKV', version=1, nLayers=70, tokens=16, inner[70] uint32, then per layer i: k_ip_i [tokens*inner_i] f32, v_ip_i [tokens*inner_i] f32. Matches native sdxl_main.cpp::SdxlIpKv::load + ImageGenerationSdxlNpuIpKv."
   }
 }