Integrate trained NumZoo LoRA into the app

- Bundle the step-1250 FLUX.2-klein-4B style LoRA (lora/numzoo_klein.safetensors,
88MB via Git LFS) — chosen for strongest style + clean multi-animal, no overfit
- Load it in _load_pipeline_cpu() after from_pretrained (module/CPU scope, outside
the ZeroGPU budget); bulletproof try/except falls back to base model
- build_prompt() now prepends the "NUMZOO" trigger and appends "no text" (the
distilled 4-step klein otherwise renders the trigger word as a literal sign)
- Add peft to requirements.txt (required by load_lora_weights)

Local MPS test: cozy NumZoo style renders correctly, no trigger text, multi-animal
works. All 10 Puppeteer checks pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitattributes

@@ -1 +1,2 @@
 training/*.jpg filter=lfs diff=lfs merge=lfs -text
+lora/*.safetensors filter=lfs diff=lfs merge=lfs -text

image_generator.py

@@ -215,8 +215,17 @@ def build_subject(animals: list[str], places: list[str]) -> str:
     return f"A cute {animal_text} {place_text}"
 
 
+# Trigger word the NumZoo LoRA learned to associate with the whole cozy aesthetic.
+# Training captions were "NUMZOO. A cute {animals} {places}, <detail>", so the live
+# prompt mirrors that prefix exactly — the trigger now carries the style, making the
+# verbose NUMZOO_STYLE suffix redundant (kept as a constant for the dataset generator).
+NUMZOO_TRIGGER = "NUMZOO"
+
+
 def build_prompt(animals: list[str], places: list[str]) -> str:
-    return f"{build_subject(animals, places)}, {NUMZOO_STYLE}"
+    # "no text" suppresses the model rendering the NUMZOO trigger word as a literal
+    # sign/title (the distilled 4-step klein is prone to this without it).
+    return f"{NUMZOO_TRIGGER}. {build_subject(animals, places)}, no text"
 
 # ---------------------------------------------------------------------------
 # Pipeline loader — two-phase, following the reference ZeroGPU pattern:
@@ -254,6 +263,20 @@ def _load_pipeline_cpu() -> None:
         torch_dtype=_dtype,
         token=hf_token,
     )
+
+    # Apply the NumZoo style LoRA (trained on FLUX.2-klein-base-4B, loads on distilled).
+    # Bundled in the repo via Git LFS. Loaded here at module/CPU scope so it stays out
+    # of the ZeroGPU 60 s budget. Never crash the import — fall back to the base model.
+    lora_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "lora", "numzoo_klein.safetensors")
+    if os.path.isfile(lora_path):
+        try:
+            _pipe.load_lora_weights(lora_path)
+            print(f"✅ NumZoo LoRA loaded from {lora_path}")
+        except Exception as e:
+            print(f"⚠️  Could not load NumZoo LoRA ({e}) — using base model")
+    else:
+        print(f"⚠️  NumZoo LoRA not found at {lora_path} — using base model")
+
     print("Pipeline loaded on CPU — ready for device placement.")
 
 

lora/numzoo_klein.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:077173e834475031559123fbdc5bed64a267aa8a3e7e389e6ac7ed1b8af8e2ae
+size 92426536

requirements.txt

@@ -1,6 +1,7 @@
 diffusers>=0.30.0
 transformers>=4.44.0
 accelerate>=0.33.0
+peft>=0.11.0
 torch>=2.2.0
 numpy<2
 sentencepiece