Adapter v12 in the DiT: 2304-slot position table, trained at the real edit geometry

The bundled text_fusion is now revision v12 (vision cosine against the native Qwen3-VL-8B encoder
0.930 -> 0.971, text unchanged at 0.951/24.2%). Two things moved: the attention-branch position table
covers 2304 slots, and the fine-tune ran at the real inference geometry (--geom 1024), so the ~2000
token condition of a two-image edit no longer falls into a zero-padded tail.

transformer/config.json text_fusion_config.max_len goes 512 -> 2304 to match the table; the model card,
the conditioning row and the card media are re-rendered with the shipped weights.

README.md

@@ -28,7 +28,7 @@ Text-to-image, character and scene editing, and transparent (RGBA) generation in
 |---|---|
 | transformer | Qwen-Image-2.1 DiT — 32 layers, 14.5 GB fp16, plus a **158M text-fusion adapter** inside |
 | text encoder | **Qwen3.5-0.8B**, 1.7 GB fp16 — upstream checkpoint re-saved to fp16, tokenizer/processor files unchanged (native: Qwen3-VL-8B, 17.5 GB) |
-| conditioning | cosine **0.94** against the native Qwen3-VL-8B encoder (text positions) |
+| conditioning | cosine **0.95** on text, **0.97** on the vision positions of edit prompts, against the native Qwen3-VL-8B encoder |
 | VAE | Qwen-Image-2.1, 16× spatial, fp32 |
 | scheduler | `FlowMatchEulerDiscreteScheduler`, plain static shift 5.0 (dynamic shifting off) |
 | resolution | `output_resolution`, 1024 by default; follows the condition image aspect ratio |
@@ -43,6 +43,11 @@ images (**Improved using Qwen**). The adapter lives *inside* the DiT as its text
 whole model is one self-contained diffusers folder and no 17.5 GB encoder is needed anywhere. The
 sampler runs a plain static shift of 5.0 instead of the original dynamic shifting.
 
+The bundled adapter is revision **v12**: its attention-branch position table covers 2304 slots and it
+was fine-tuned at the real inference geometry (~2000-token conditions at 1024 px), so reference images
+keep their positions instead of falling into a zero-padded tail — the vision cosine against the native
+encoder moved 0.93 → 0.97, text stayed at 0.95.
+
 ### Examples
 
 Every image below is generated by this pipeline with 30 steps at 1024 px.
@@ -64,7 +69,8 @@ pose, clothing and scene; the identity is copied from `<image2>`)
 |---|---|
 | ![edit swap](media/edit_swap.jpg) | ![edit char](media/edit_char.jpg) |
 
-**Edit — three condition images** (subject from `<image1>`, scene from `<image2>`, lighting from `<image3>`)
+**Edit — three condition images** (target and composition from `<image1>`, the person from `<image2>`,
+colour and lighting from `<image3>`)
 
 ![edit three](media/edit_three.jpg)
 
@@ -170,6 +176,8 @@ would not find the DiT class.
 * **English only** — that is all the adapter was trained and tested on; other languages drift.
 * **Numerals on signage** come out wrong: "OPEN 24 HOURS" renders as "OPEN **26** HOURS" on every
   seed tried. Words are fine. ![numbers](media/limit_numbers.jpg)
+* **Non-photo references transfer less faithfully** than photographic ones: the adapter imitates the
+  native encoder, so its ceiling is the native encoder's ceiling.
 * Batch size >1 at 1024 px peaks near 28 GB; one prompt per call is the safe mode.
 
 ### NOTICE

transformer/config.json

@@ -26,7 +26,7 @@
     "attention": 2,
     "attn_dim": 1024,
     "attn_heads": 8,
-    "max_len": 512,
+    "max_len": 2304,
     "mixer": 2,
     "mixer_heads": 8,
     "mixer_ffn": 2,

transformer/diffusion_pytorch_model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a71f3c4eca68c67df8c8fd288cc1daf27ac0a89d6eee9f0ab7aa2ec1031adbe4
-size 10284159452
+oid sha256:3e706b011865bbcf5e36c9a62e189549179de9fcbf8c886afc110ce2b30992bc
+size 10287829484