Upload 8 files

STYLE_TRAINING.md

@@ -1,10 +1,19 @@
-# Custom Style LoRA
+# Real Style LoRA workflow
 
-Use only images you own or are licensed to train on.
+1. Keep the 10 reference images in `training_data/`.
+2. Run `python prepare_dataset.py`.
+3. Run `python train_style_lora.py` on a CUDA GPU.
+4. The official Diffusers SDXL DreamBooth-LoRA trainer writes safetensors weights into `weights/`.
+5. Start `app.py`. The app auto-loads the local LoRA if `weights/pytorch_lora_weights.safetensors` exists.
 
-Put training images in `training_data/`. `train_style_lora.py` validates them and
-creates a manifest. Use a current Diffusers LoRA training workflow compatible
-with your GPU to actually train the weights.
+### Environment tuning
+- `MAX_TRAIN_STEPS=1200`
+- `LEARNING_RATE=1e-4`
+- `LORA_RANK=16`
+- `GRAD_ACCUM=4`
+- `TRAIN_RESOLUTION=1024`
+- `TRAIN_BATCH_SIZE=1`
+- `MODEL_ID=stabilityai/stable-diffusion-xl-base-1.0`
+- `HF_TOKEN=<your token>` if the Space needs authenticated model access
 
-The LoRA teaches the target visual style/character patterns. The expression
-engine and reference-image adapter handle expression/pose control separately.
+The 10 images are a small dataset. They can teach a recognizable style, but a larger set with more poses/characters generally gives a more robust style LoRA.

ZEROGPU_TRAINING.md

@@ -9,3 +9,6 @@ The Space UI must connect a Gradio button to `train_style_lora.train_chunk`.
 Run the button repeatedly until the status reports completion.
 
 A full SDXL LoRA run is not guaranteed to fit into a single ZeroGPU call or daily quota.
+
+
+Note: the trainer downloads the maintained SDXL DreamBooth LoRA script from the Diffusers examples/dreambooth path.

app.py

@@ -1,7 +1,9 @@
 import os, random, gc, json, re
 from pathlib import Path
+import spaces  # must be imported before torch for ZeroGPU startup detection
 import gradio as gr
 import torch
+from train_style_lora import train_chunk
 from PIL import Image
 from rembg import remove, new_session
 from diffusers import (
@@ -18,21 +20,21 @@ from transformers import CLIPVisionModelWithProjection
 # ============================================================
 
 MODEL_ID = os.getenv("MODEL_ID", "stabilityai/stable-diffusion-xl-base-1.0")
-LORA_REPO = os.getenv("LORA_REPO", "")
-LORA_WEIGHT = os.getenv("LORA_WEIGHT", "")
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
+LORA_REPO = os.getenv("LORA_REPO", str(Path(__file__).parent / "weights"))
+LORA_WEIGHT = os.getenv("LORA_WEIGHT", "offex3dstyle.safetensors")
 
 # Fixed identity/style contract. Keep this stable across generations.
 STYLE_LOCK = """
-high-end stylized 3D animated character asset,
+offex3dstyle, high-end stylized 3D animated character asset,
 premium polished character design, expressive stylized face,
 clean sculpted facial forms, large readable eyes,
 clear eyebrows and mouth shapes, detailed clothing,
 clean hands and fingers, strong silhouette,
 full body from head to toe, entire feet visible,
 entire arms and hands visible, centered character,
-consistent proportions, professional production character sheet quality,
+consistent proportions, professional production character asset quality,
+ONE AND ONLY ONE CHARACTER, a single person/character only,
+no character sheet, no turnaround sheet, no multiple views, no collage, no thumbnails, no duplicate character,
 isolated character, transparent-background asset look,
 neutral balanced studio lighting, clean edges, sharp details,
 no environment, no props unless requested, no text, no logo, no watermark
@@ -40,7 +42,9 @@ no environment, no props unless requested, no text, no logo, no watermark
 
 NEGATIVE = """
 cropped, cut off head, cut off feet, half body, close-up,
-duplicate character, two characters, extra arms, extra legs,
+character sheet, character turnaround, turnaround sheet, model sheet, expression sheet, multiple views, multiple poses, collage, grid, contact sheet, headshot, bust portrait, portrait crop, close-up,
+duplicate character, two characters, three characters, multiple characters, extra character, repeated character,
+extra arms, extra legs,
 extra fingers, fused fingers, missing fingers, deformed hands,
 bad anatomy, malformed face, asymmetrical eyes, floating limbs,
 background scene, landscape, room, street, text, letters, logo,
@@ -190,25 +194,32 @@ _lora_loaded = False
 
 def load_models():
     global _pipe, _rmbg, _lora_loaded
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    dtype = torch.float16 if device == "cuda" else torch.float32
     if _pipe is None:
+        if device != "cuda":
+            raise RuntimeError("GPU allocation is required for generation. Please retry while the ZeroGPU function is running.")
         _pipe = AutoPipelineForText2Image.from_pretrained(
-            MODEL_ID, torch_dtype=DTYPE, use_safetensors=True
+            MODEL_ID, torch_dtype=dtype, use_safetensors=True
         )
         if LORA_REPO and LORA_WEIGHT:
             try:
-                _pipe.load_lora_weights(LORA_REPO, weight_name=LORA_WEIGHT, adapter_name="character_style")
+                # Load the locally trained style LoRA when the weight file exists.
+                weight_path = Path(LORA_REPO) / LORA_WEIGHT
+                if not weight_path.exists() and Path(LORA_REPO).is_file():
+                    weight_path = Path(LORA_REPO)
+                if not weight_path.exists():
+                    raise FileNotFoundError(f"Style LoRA not found: {weight_path}")
+                _pipe.load_lora_weights(str(weight_path.parent), weight_name=weight_path.name, adapter_name="character_style")
                 _pipe.set_adapters(["character_style"], adapter_weights=[float(os.getenv("LORA_SCALE","0.75"))])
                 _lora_loaded = True
             except Exception as e:
                 print("LoRA load failed:", e)
 
-        if DEVICE == "cuda":
-            _pipe = _pipe.to("cuda")
-            try: _pipe.enable_xformers_memory_efficient_attention()
-            except Exception: pass
-            _pipe.enable_attention_slicing()
-        else:
-            _pipe = _pipe.to("cpu")
+        _pipe = _pipe.to(device)
+        try: _pipe.enable_xformers_memory_efficient_attention()
+        except Exception: pass
+        _pipe.enable_attention_slicing()
     if _rmbg is None:
         _rmbg = new_session("u2net_human_seg")
     return _pipe, _rmbg
@@ -242,16 +253,21 @@ EXTRA DIRECTION:
 {extra or "None"}
 
 OUTPUT:
-single full-body character, centered, head-to-toe, isolated asset
+ONE SINGLE FULL-BODY CHARACTER ONLY.
+This is a single-character asset, not a character sheet and not a portrait collection.
+Exactly one single character in the entire image, centered and occupying most of the canvas.
+Full-body standing pose from top of head to soles of both feet, one person only, large enough to fill the vertical frame, both feet visible, both arms visible, isolated on a clean transparent-style background.
+Do NOT show a headshot, bust, portrait crop, front/back/side turnaround, multiple views, additional poses, smaller copies, duplicate copies, thumbnails, character sheet, expression sheet, collage, grid, contact sheet, or any second character.
 """.strip()
 
-def generate(character, expression, consistency, reference, strength, extra, steps, guidance, seed):
+@spaces.GPU(duration=120, size="large")
+def generate_character(character, expression, consistency, reference, strength, extra, steps, guidance, seed):
     if not character or not character.strip():
         raise gr.Error("Character description डालना जरूरी है।")
     pipe, rmbg = load_models()
 
     seed = random.randint(0, 2147483647) if seed is None or int(seed) < 0 else int(seed)
-    generator = torch.Generator(device=DEVICE).manual_seed(seed)
+    generator = torch.Generator(device="cuda").manual_seed(seed)
     prompt = make_prompt(character, expression, consistency, reference, extra)
 
     # IP-Adapter is the reference-image pathway when the installed diffusers
@@ -259,7 +275,7 @@ def generate(character, expression, consistency, reference, strength, extra, ste
     # combinations differ between Spaces.
     kwargs = dict(
         prompt=prompt, negative_prompt=NEGATIVE,
-        width=1024, height=1024,
+        width=832, height=1216,
         num_inference_steps=int(steps),
         guidance_scale=float(guidance),
         generator=generator
@@ -306,7 +322,7 @@ button.primary{background:linear-gradient(100deg,#7357ff,#22d4ff)!important;bord
 .output{border-radius:22px!important;min-height:650px!important}
 """
 
-with gr.Blocks(title="AI Character Studio ULTRA", css=CSS, theme=gr.themes.Base()) as demo:
+with gr.Blocks(title="AI Character Studio ULTRA") as demo:
     gr.HTML("""
     <div class="hero">
       <div style="color:#a9b3c7;font-size:12px;letter-spacing:2px;margin-bottom:8px">AI CHARACTER STUDIO • ULTRA</div>
@@ -353,8 +369,10 @@ with gr.Blocks(title="AI Character Studio ULTRA", css=CSS, theme=gr.themes.Base(
                 guidance = gr.Slider(4,10,7,step=.5,label="Guidance")
 
             seed = gr.Number(value=-1, precision=0, label="Seed (-1 = random)")
-            generate = gr.Button("✨ GENERATE HIGH-QUALITY CHARACTER", variant="primary", elem_id="generate")
+            generate_btn = gr.Button("✨ GENERATE HIGH-QUALITY CHARACTER", variant="primary", elem_id="generate")
             clear = gr.Button("Clear GPU Cache")
+            train_lora = gr.Button("🔥 TRAIN LORA CHUNK (ZeroGPU)", variant="primary")
+            train_status = gr.Textbox(label="LoRA Training Status", interactive=False, lines=3)
             status = gr.Textbox(label="System Status", interactive=False)
 
         with gr.Column(scale=7):
@@ -362,12 +380,13 @@ with gr.Blocks(title="AI Character Studio ULTRA", css=CSS, theme=gr.themes.Base(
             used_seed = gr.Number(label="Seed Used", interactive=False)
             prompt_view = gr.Textbox(label="Resolved Prompt", lines=14, interactive=False)
 
-    generate.click(
-        generate,
+    generate_btn.click(
+        generate_character,
         [character, expression, consistency, reference, strength, extra, steps, guidance, seed],
         [output, used_seed, prompt_view]
     )
     clear.click(clear_cache, outputs=status)
+    train_lora.click(train_chunk, outputs=train_status)
 
 if __name__ == "__main__":
-    demo.queue(max_size=20).launch()
+    demo.queue(max_size=20).launch(css=CSS, theme=gr.themes.Base())

requirements.txt

@@ -1,10 +1,14 @@
-gradio>=5.38.0
-diffusers>=0.35.1
-transformers>=4.53.0
-accelerate>=1.8.1
-safetensors>=0.5.3
-Pillow>=11.0.0
-rembg[gpu]>=2.0.67
-onnxruntime-gpu>=1.22.0
-numpy<2
-peft>=0.15.0
+# SDXL generation + real LoRA training
+accelerate>=1.2.0
+diffusers>=0.31.0
+transformers>=4.46.0
+peft>=0.13.0
+safetensors>=0.4.5
+huggingface_hub>=0.26.0
+bitsandbytes>=0.45.0
+gradio>=5.0.0
+Pillow>=10.0.0
+rembg>=2.0.59
+onnxruntime-gpu>=1.19.0; platform_system == "Linux"
+spaces>=0.36.0
+torchvision>=0.20.0

train_style_lora.py

@@ -31,7 +31,7 @@ HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HF_HUB_TOKEN", "")
 TRAIN_SCRIPT_URL = os.getenv(
     "DIFFUSERS_TRAIN_SCRIPT_URL",
     "https://raw.githubusercontent.com/huggingface/diffusers/main/"
-    "examples/advanced_diffusion_training/train_dreambooth_lora_sdxl.py",
+    "examples/dreambooth/train_dreambooth_lora_sdxl.py",
 )
 
 # Keep each ZeroGPU request bounded. Run Train again to continue the next chunk.