fix: use ref_images + zero_cond_t for proper editing

app.py

@@ -111,6 +111,7 @@ def load_engine(model_name=None):
         vae_path=vae_path,
         offload_params_to_cpu=True,
         diffusion_flash_attn=True,
+        qwen_image_zero_cond_t=True,
         n_threads=N_THREADS,
         verbose=True,
     )
@@ -174,7 +175,7 @@ def safe_load_image(path, max_px=MAX_INPUT_PX, crop_ratio=None):
             print(f"[gen] Center-cropped to {img.size[0]}x{img.size[1]} for {target_w}:{target_h} ratio", flush=True)
     return img
 
-def generate(prompt, negative_prompt, init_image, model_choice, aspect_ratio, steps, cfg_scale, guidance, strength, seed):
+def generate(prompt, negative_prompt, init_image, model_choice, aspect_ratio, steps, cfg_scale, guidance, seed):
     gc.collect()
     print(f"\n{'='*60}", flush=True)
     print(f"[gen] START {time.strftime('%H:%M:%S')}", flush=True)
@@ -213,8 +214,7 @@ def generate(prompt, negative_prompt, init_image, model_choice, aspect_ratio, st
     )
 
     if pil_input is not None:
-        kwargs["init_image"] = pil_input
-        kwargs["strength"] = float(strength)
+        kwargs["ref_images"] = [pil_input]
 
     mode = "edit" if pil_input else "txt2img"
     print(f"[gen] {mode} {w}x{h} steps={steps} cfg={cfg_scale} guidance={guidance} seed={seed}", flush=True)
@@ -294,11 +294,10 @@ def cli_main():
     parser.add_argument("--steps", type=int, default=4)
     parser.add_argument("--cfg", type=float, default=2.5)
     parser.add_argument("--guidance", type=float, default=3.0)
-    parser.add_argument("--strength", type=float, default=0.3)
     parser.add_argument("--seed", type=int, default=-1)
     args = parser.parse_args()
 
-    for img, status in generate(args.prompt, args.negative, args.init_image, args.model, args.aspect, args.steps, args.cfg, args.guidance, args.strength, args.seed):
+    for img, status in generate(args.prompt, args.negative, args.init_image, args.model, args.aspect, args.steps, args.cfg, args.guidance, args.seed):
         if img:
             img.save(args.output)
             print(f"Saved: {args.output} ({status})")
@@ -328,8 +327,6 @@ def gradio_main():
                     steps = gr.Slider(1, 50, value=4, step=1, label="Steps", scale=1)
                     cfg_scale = gr.Slider(1.0, 7.0, value=2.5, step=0.5, label="CFG", scale=1)
                     guidance = gr.Slider(1.0, 10.0, value=3.0, step=0.5, label="Guidance", scale=1)
-                with gr.Row():
-                    strength = gr.Slider(0.1, 1.0, value=0.3, step=0.05, label="Strength (edit)", scale=1)
                     seed = gr.Number(value=-1, label="Seed", precision=0, scale=1)
             with gr.Column(variant="panel", scale=1, min_width=280):
                 output_image = gr.Image(label="Output", type="pil", height=380)
@@ -343,7 +340,7 @@ def gradio_main():
 
         gen_btn.click(
             fn=generate,
-            inputs=[prompt, negative_prompt, init_image, model_choice, aspect_ratio, steps, cfg_scale, guidance, strength, seed],
+            inputs=[prompt, negative_prompt, init_image, model_choice, aspect_ratio, steps, cfg_scale, guidance, seed],
             outputs=[output_image, status_text],
             api_name="infer", concurrency_limit=1,
         )