Upgrade to FLUX.2-klein-4B (Apache, 4B, sub-second) via Flux2KleinPipeline (#2)

- Upgrade to FLUX.2-klein-4B (Apache, 4B, sub-second) via Flux2KleinPipeline (9a693860e8b78284eadf391eeaec4bcc8e51d269)
- FLUX.2-klein needs diffusers from git (Flux2KleinPipeline) + Qwen3 text encoder deps (636d2166310638ff70a796fb5c917f39a689c3d6)

app.py

@@ -4,19 +4,23 @@ import gradio as gr
 import numpy as np
 import random
 import torch
-from diffusers import DiffusionPipeline
+from diffusers import Flux2KleinPipeline
 
 # ---------------------------------------------------------------------------
-# Model
+# Model: FLUX.2 [klein] 4B
+#   - Apache-2.0, 4B params, BFL's fastest small model (sub-second, ~13GB VRAM)
+#   - Unified text-to-image + multi-reference editing
+#   - Released Jan 2026 (current BFL small-model generation)
 # ---------------------------------------------------------------------------
-# FLUX.1-schnell: Apache-2.0, 1-4 step distilled text-to-image. Fast on ZeroGPU.
-MODEL_REPO_ID = "black-forest-labs/FLUX.1-schnell"
+MODEL_REPO_ID = "black-forest-labs/FLUX.2-klein-4B"
 
 dtype = torch.bfloat16
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# Load at module level on cuda (ZeroGPU CUDA-emulation makes this safe outside @spaces.GPU)
-pipe = DiffusionPipeline.from_pretrained(MODEL_REPO_ID, torch_dtype=dtype).to(device)
+# Load on cuda at module level. (No enable_model_cpu_offload() on ZeroGPU —
+# the GPU is only attached inside @spaces.GPU; module-level cuda uses the
+# ZeroGPU CUDA-emulation, and offload would conflict.)
+pipe = Flux2KleinPipeline.from_pretrained(MODEL_REPO_ID, torch_dtype=dtype).to(device)
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 2048
@@ -30,6 +34,7 @@ def infer(
     width,
     height,
     num_inference_steps,
+    guidance_scale,
     progress=gr.Progress(track_tqdm=True),
 ):
     if randomize_seed:
@@ -40,8 +45,7 @@ def infer(
         width=width,
         height=height,
         num_inference_steps=num_inference_steps,
-        guidance_scale=0.0,        # schnell is guidance-distilled
-        max_sequence_length=256,
+        guidance_scale=guidance_scale,
         generator=generator,
     ).images[0]
     return image, seed
@@ -49,8 +53,8 @@ def infer(
 
 examples = [
     "A magical city at twilight, glowing windows, storybook illustration, warm light",
+    "A cat holding a sign that says hello world",
     "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "A delicious ceviche cheesecake slice",
 ]
 
 css = """
@@ -59,7 +63,7 @@ css = """
 
 with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown("# 🖼️ NEXUS Visual Weaver — FLUX.1-schnell")
+        gr.Markdown("# 🖼️ NEXUS Visual Weaver — FLUX.2 [klein] 4B")
         with gr.Row():
             prompt = gr.Text(
                 label="Prompt",
@@ -82,13 +86,16 @@ with gr.Blocks(css=css) as demo:
                 num_inference_steps = gr.Slider(
                     label="Inference steps", minimum=1, maximum=8, step=1, value=4
                 )
+                guidance_scale = gr.Slider(
+                    label="Guidance scale", minimum=0.0, maximum=5.0, step=0.1, value=1.0
+                )
 
         gr.Examples(examples=examples, inputs=[prompt])
 
     gr.on(
         triggers=[run_button.click, prompt.submit],
         fn=infer,
-        inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps],
+        inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps, guidance_scale],
         outputs=[result, seed],
     )
 

requirements.txt

@@ -1,6 +1,6 @@
 spaces
 accelerate
-diffusers
+git+https://github.com/huggingface/diffusers.git
 transformers
 sentencepiece
 torch