flux2 latent preview via Modular Diffusers streaming API (pipe.stream, diffusers PR #14159)

README.md

@@ -1,13 +1,20 @@
 ---
 title: Flux2 Latent Preview
-emoji: 👀
-colorFrom: pink
-colorTo: yellow
+emoji: 🌊
+colorFrom: yellow
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.24.0
-python_version: '3.13'
+sdk_version: 6.5.1
+python_version: '3.12'
 app_file: app.py
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Live latent preview for FLUX.2 [dev] using the Modular Diffusers streaming API (`pipe.stream()`)
+from [huggingface/diffusers#14159](https://github.com/huggingface/diffusers/pull/14159).
+
+Unlike the [flux-latent-preview](https://huggingface.co/spaces/diffusers-internal-dev/flux-latent-preview)
+space (which inserts a custom preview block into the denoise loop and drains a queue from a worker
+thread), this app just iterates `pipe.stream()` — the pipeline yields its live state after every
+denoising step — and renders each state with a preview pipeline assembled from flux2's own
+unpack + decode blocks, sharing the main pipeline's VAE.

app.py

@@ -0,0 +1,163 @@
+import random
+
+import gradio as gr
+import numpy as np
+import spaces
+import torch
+
+from diffusers import ModularPipeline
+from diffusers.modular_pipelines import SequentialPipelineBlocks
+from diffusers.modular_pipelines.flux2.decoders import Flux2DecodeStep, Flux2UnpackLatentsStep
+
+
+pipe = ModularPipeline.from_pretrained("diffusers/flux2-bnb-4bit-modular")
+pipe.load_components(torch_dtype=torch.bfloat16, device_map="cuda")
+
+# The live-preview decoder is a second pipeline assembled from flux2's own blocks,
+# sharing the main pipeline's VAE — no custom block, no queue, no thread.
+preview = SequentialPipelineBlocks.from_blocks_dict(
+    {"unpack": Flux2UnpackLatentsStep(), "decode": Flux2DecodeStep()}
+).init_pipeline()
+preview.update_components(vae=pipe.vae)
+
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 2048
+
+
+@spaces.GPU(duration=120)
+def infer(
+    prompt,
+    seed=42,
+    randomize_seed=False,
+    width=1024,
+    height=1024,
+    guidance_scale=4.0,
+    num_inference_steps=28,
+    progress=gr.Progress(track_tqdm=True),
+):
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    generator = torch.Generator().manual_seed(seed)
+
+    # `pipe.stream()` yields an event with the live pipeline state after every denoising step
+    stream = pipe.stream(
+        prompt=prompt,
+        guidance_scale=guidance_scale,
+        num_inference_steps=num_inference_steps,
+        width=width,
+        height=height,
+        generator=generator,
+    )
+    while True:
+        try:
+            event = next(stream)
+        except StopIteration as e:
+            state = e.value
+            break
+        # flow matching: after step i the latents sit at sigmas[i + 1]; project to the predicted
+        # clean image x0 = x_t - sigma * v so the preview shows the image forming, not noise
+        latents = event.state.get("latents")
+        sigma = pipe.scheduler.sigmas[event.loop_kwargs["i"] + 1].to(latents.device, latents.dtype)
+        x0 = latents - sigma * event.state.get("noise_pred")
+        image = preview(
+            latents=x0,
+            latent_ids=event.state.get("latent_ids"),
+            output="images",
+        )[0]
+        yield image, seed
+
+    # the final image decoded by the pipeline's own decode step
+    yield state.get("images")[0], seed
+
+
+examples = [
+    "a tiny astronaut hatching from an egg on the moon",
+    "a cat holding a sign that says hello world",
+    "an anime illustration of a wiener schnitzel",
+]
+
+css = """
+#col-container {
+    margin: 0 auto;
+    max-width: 520px;
+}
+"""
+
+with gr.Blocks() as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(
+            """# FLUX.2 [dev] — Live Preview with Modular Diffusers
+Live latent preview powered by `pipe.stream()`: the pipeline yields its live state after every
+denoising step, and a preview pipeline built from flux2's own unpack + decode blocks renders it.
+No custom blocks, queues, or threads — see [huggingface/diffusers#14159](https://github.com/huggingface/diffusers/pull/14159).
+"""
+        )
+
+        with gr.Row():
+            prompt = gr.Text(
+                label="Prompt",
+                show_label=False,
+                max_lines=1,
+                placeholder="Enter your prompt",
+                container=False,
+            )
+
+            run_button = gr.Button("Run", scale=0)
+
+        result = gr.Image(label="Result", show_label=False)
+
+        with gr.Accordion("Advanced Settings", open=False):
+            seed = gr.Slider(
+                label="Seed",
+                minimum=0,
+                maximum=MAX_SEED,
+                step=1,
+                value=0,
+            )
+
+            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+
+            with gr.Row():
+                width = gr.Slider(
+                    label="Width",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+
+                height = gr.Slider(
+                    label="Height",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+
+            with gr.Row():
+                guidance_scale = gr.Slider(
+                    label="Guidance Scale",
+                    minimum=1,
+                    maximum=15,
+                    step=0.1,
+                    value=4.0,
+                )
+
+                num_inference_steps = gr.Slider(
+                    label="Number of inference steps",
+                    minimum=1,
+                    maximum=50,
+                    step=1,
+                    value=28,
+                )
+
+        gr.Examples(examples=examples, fn=infer, inputs=[prompt], outputs=[result, seed], cache_examples=False)
+
+    gr.on(
+        triggers=[run_button.click, prompt.submit],
+        fn=infer,
+        inputs=[prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
+        outputs=[result, seed],
+    )
+
+demo.launch(css=css)

requirements.txt

@@ -0,0 +1,5 @@
+accelerate
+git+https://github.com/huggingface/diffusers.git@refs/pull/14159/head
+torch
+transformers
+bitsandbytes