update

python_server/model/Inpaint/flux_fill/loader.py

@@ -7,6 +7,32 @@ import numpy as np
 from PIL import Image
 
 
+def _align_spatial_dim_for_flux_fill(n: int, multiple: int = 16) -> int:
+    """FLUX Fill 要求 H/W 为 (vae_scale_factor*2) 的倍数，常见为 16。"""
+    n = int(max(n, multiple))
+    return n - (n % multiple)
+
+
+def _flux_clip_prompt_and_t5_full(pipe: Any, text_full: str) -> tuple[str, str]:
+    """
+    CLIP 仅约 77 token：用与管线一致的 tokenizer 截断得到 prompt（池化向量）。
+    完整文本走 prompt_2 → T5（max_sequence_length 可达 512），长中文不再整段喂给 CLIP。
+    """
+    text_full = (text_full or "").strip()
+    if not text_full:
+        d = "Chinese ink painting background, natural texture, seamless fill, no figures"
+        return d, d
+    tok = getattr(pipe, "tokenizer", None)
+    if tok is None:
+        return text_full[:120], text_full
+    ml = int(getattr(tok, "model_max_length", 77) or 77)
+    enc = tok([text_full], padding="max_length", max_length=ml, truncation=True, return_tensors="pt")
+    clip_s = tok.batch_decode(enc.input_ids, skip_special_tokens=True)[0].strip()
+    if not clip_s:
+        clip_s = text_full[:80]
+    return clip_s, text_full
+
+
 def make_flux_fill_predictor(
     model_id: str,
     device: str | None = None,
@@ -69,8 +95,8 @@ def make_flux_fill_predictor(
             scale = max_side / float(max(orig_w, orig_h))
             run_w = int(round(orig_w * scale))
             run_h = int(round(orig_h * scale))
-        run_w = max(8, run_w - (run_w % 8))
-        run_h = max(8, run_h - (run_h % 8))
+        run_w = _align_spatial_dim_for_flux_fill(run_w)
+        run_h = _align_spatial_dim_for_flux_fill(run_h)
 
         if (run_w, run_h) != (orig_w, orig_h):
             image_run = image.resize((run_w, run_h), resample=Image.BICUBIC)
@@ -79,9 +105,10 @@ def make_flux_fill_predictor(
             image_run = image
             mask_run = mask
 
-        p = (prompt or "").strip()
-        if not p:
-            p = "Chinese ink painting background, natural texture, seamless fill, no figures"
+        p_full = (prompt or "").strip()
+        if not p_full:
+            p_full = "Chinese ink painting background, natural texture, seamless fill, no figures"
+        p_clip, p_t5 = _flux_clip_prompt_and_t5_full(pipe, p_full)
 
         st = float(strength)
         st = min(1.0, max(0.05, st))
@@ -93,7 +120,8 @@ def make_flux_fill_predictor(
                 pass
 
         out = pipe(
-            prompt=p,
+            prompt=p_clip,
+            prompt_2=p_t5,
             image=image_run,
             mask_image=mask_run,
             height=run_h,