Switch to Space node calling ideogram-ai/ideogram4 /generate

A kind:'space' node authenticates with each visitor's HF OAuth token
(gradio@6.19.0 call_space passes the resolved token to gradio_client),
so no server-side key is needed and the fal bound-function workaround
is unnecessary. The /generate endpoint already exposes all controls
(prompt, mode, upsampler, width, height, seed, randomize_seed) and
returns (image, seed, caption), mapped to input/output ports by order.

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

README.md

@@ -17,13 +17,16 @@ hf_oauth_scopes:
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
-## Secrets
+## How it works
 
-This Space calls Ideogram 4 through the **fal** inference provider, which
-exposes more controls than the stock text-to-image call. Set at least one of:
+This is a Gradio **Workflow** app. The canvas calls the official
+[Ideogram 4 Space](https://huggingface.co/spaces/ideogram-ai/ideogram4)
+(`ideogram-ai/ideogram4`, endpoint `/generate`) as a node, exposing all of its
+inputs — `prompt`, `mode` (speed ↔ quality), `upsampler`, `width`, `height`,
+`seed`, `randomize_seed` — as ports you can wire and tweak on the canvas. The
+node returns the image, the seed used, and the (upsampled) caption.
 
-- `FAL_KEY` — a fal API key (fal's recommended auth). Get one at https://fal.com/dashboard/keys
-- `HF_TOKEN` — a Hugging Face token with Inference access (used as a fallback
-  and to let the Workflow canvas authenticate each visitor via OAuth).
-
-If neither is set the generator node will fail with an auth error at run time.
+No server-side secrets are required: each visitor authenticates with their own
+Hugging Face OAuth token (granted via `hf_oauth` below), so they run the
+pipeline on their own access. The Space owner gets write access to edit and save
+the workflow; visitors get a read-only canvas they can still run.

app.py

@@ -1,84 +1,15 @@
-import os
-from typing import Optional
-
 import gradio as gr
-from huggingface_hub import InferenceClient
-
-
-# Fal's Ideogram 4 endpoint exposes more controls than the stock text-to-image
-# call. gradio's built-in `kind: "model"` node only forwards the prompt to
-# `InferenceClient.text_to_image(prompt)` (verified in gradio@6.19.0, where the
-# `text-to-image` branch is literally `client.text_to_image(a0)` — no
-# `extra_body`), so fal-specific fields would be silently dropped. We therefore
-# bind this Python function as a node and forward the extra controls ourselves
-# via `extra_body`, which is exactly how huggingface_hub passes provider-specific
-# parameters to fal.
-def ideogram_v4(
-    prompt: str,
-    rendering_speed: str = "BALANCED",
-    image_size: str = "square_hd",
-    expansion_model: str = "Medium",
-    acceleration: str = "none",
-    num_images: int = 1,
-    seed: Optional[int] = None,
-    output_format: str = "jpeg",
-    enable_safety_checker: bool = True,
-):
-    """Generate image(s) with Ideogram 4 via the fal inference provider.
-
-    Drag this node onto the canvas and wire the `prompt` port from a text
-    reference; tweak the other ports for control over quality, aspect ratio,
-    prompt expansion, count, seed and output format.
-    """
-    client = InferenceClient(
-        provider="fal-ai",
-        # Prefer a fal key when present (fal's recommended auth), else fall back
-        # to the HF token (HF routes fal traffic through the visitor's HF
-        # inference entitlement on the Space).
-        token=os.environ.get("FAL_KEY") or os.environ.get("HF_TOKEN") or None,
-    )
-
-    # fal accepts image_size as either a preset string ("square_hd", "landscape_16_9",
-    # ...) or a {"width", "height"} object. We pass the string through unchanged
-    # unless the user gave "WxH", which we convert to the object form.
-    size_field = image_size
-    if image_size and "x" in image_size and image_size.replace("x", "").isdigit():
-        w, h = (int(x) for x in image_size.split("x"))
-        size_field = {"width": w, "height": h}
-
-    extra_body = {
-        "rendering_speed": rendering_speed,
-        "image_size": size_field,
-        "expansion_model": expansion_model,
-        "acceleration": acceleration,
-        "num_images": max(1, int(num_images or 1)),
-        "enable_safety_checker": bool(enable_safety_checker),
-        "output_format": output_format,
-    }
-    if seed is not None:
-        extra_body["seed"] = int(seed)
-
-    images = client.text_to_image(
-        prompt or "",
-        model="ideogram-ai/ideogram-4-fp8",
-        extra_body=extra_body,
-    )
-
-    # text_to_image returns a single PIL image; fal's `num_images` > 1 is honored
-    # by the provider, but the HF helper only surfaces one image per call, so we
-    # always return a one-element gallery list. Return as a list-of-image-dicts
-    # so the canvas `gallery` output port renders it correctly.
-    if isinstance(images, (list, tuple)):
-        items = list(images)
-    else:
-        items = [images]
-    return items
 
 
-# Load the pre-wired topology (Prompt -> Ideogram 4 (fal) -> Output Gallery)
-# from workflow.json and expose the generator as a callable node on the canvas
-# so it can also be re-added / rewired after edits.
-gr.Workflow(
-    graph="workflow.json",
-    bind={"Ideogram 4 (fal)": ideogram_v4},
-).launch()
+# This Workflow calls the official Ideogram 4 Space
+# (https://huggingface.co/spaces/ideogram-ai/ideogram4) as a `kind: "space"`
+# operator node. Gradio's Workflow runner authenticates each Space node with the
+# visitor's own Hugging Face OAuth token (gradio@6.19.0 call_space passes the
+# resolved token to `gradio_client.Client(...)`), so no server-side API key is
+# required — visitors run the pipeline on their own inference entitlement.
+#
+# The /generate endpoint accepts 7 positional parameters (prompt, mode,
+# upsampler, width, height, seed, randomize_seed) and returns a 3-tuple of
+# (image, seed, caption). The node in workflow.json declares an input port per
+# parameter and an output port per returned value, in that exact order.
+gr.Workflow(graph="workflow.json").launch()

workflow.json

@@ -1,6 +1,6 @@
 {
   "schema_version": "2",
-  "name": "Ideogram 4 (fal)",
+  "name": "Ideogram 4 Pipeline",
   "references": [
     {
       "id": "ref_prompt",
@@ -10,80 +10,121 @@
       "inputs": [{ "id": "in", "label": "Text", "type": "text" }],
       "outputs": [{ "id": "out", "label": "Text", "type": "text" }],
       "x": 60,
-      "y": 160,
-      "width": 220,
+      "y": 180,
+      "width": 240,
       "height": 124,
-      "data": { "value": "a serene mountain lake at sunrise, cinematic, highly detailed" }
+      "data": { "value": "a ginger cat wearing a tiny wizard hat reading a spellbook" }
     }
   ],
   "operators": [
     {
-      "id": "fn_Ideogram 4 (fal)",
-      "source": "fn",
-      "fn": "Ideogram 4 (fal)",
-      "kind": "transform",
-      "label": "Ideogram 4 (fal)",
-      "x": 360,
-      "y": 80,
-      "width": 280,
-      "height": 460,
+      "id": "op_ideogram4",
+      "label": "Ideogram 4",
+      "role": "operator",
+      "kind": "space",
+      "space_id": "ideogram-ai/ideogram4",
+      "endpoint": "/generate",
       "inputs": [
         { "id": "in_prompt", "label": "prompt", "type": "text", "required": true },
-        { "id": "in_rendering_speed", "label": "rendering_speed", "type": "text" },
-        { "id": "in_image_size", "label": "image_size", "type": "text" },
-        { "id": "in_expansion_model", "label": "expansion_model", "type": "text" },
-        { "id": "in_acceleration", "label": "acceleration", "type": "text" },
-        { "id": "in_num_images", "label": "num_images", "type": "number" },
+        { "id": "in_mode", "label": "mode", "type": "text" },
+        { "id": "in_upsampler", "label": "upsampler", "type": "text" },
+        { "id": "in_width", "label": "width", "type": "number" },
+        { "id": "in_height", "label": "height", "type": "number" },
         { "id": "in_seed", "label": "seed", "type": "number" },
-        { "id": "in_output_format", "label": "output_format", "type": "text" },
-        { "id": "in_enable_safety_checker", "label": "enable_safety_checker", "type": "boolean" }
+        { "id": "in_randomize_seed", "label": "randomize_seed", "type": "boolean" }
       ],
       "outputs": [
-        { "id": "out_0", "label": "gallery", "type": "gallery" }
+        { "id": "out_image", "label": "Image", "type": "image", "output_index": 0 },
+        { "id": "out_seed", "label": "Seed", "type": "number", "output_index": 1 },
+        { "id": "out_caption", "label": "Caption", "type": "json", "output_index": 2 }
       ],
+      "x": 380,
+      "y": 80,
+      "width": 280,
+      "height": 460,
       "data": {
-        "in_rendering_speed": "BALANCED",
-        "in_image_size": "square_hd",
-        "in_expansion_model": "Medium",
-        "in_acceleration": "none",
-        "in_num_images": 1,
-        "in_seed": null,
-        "in_output_format": "jpeg",
-        "in_enable_safety_checker": true
+        "in_mode": "Default · 20 steps",
+        "in_upsampler": "Ideogram (remote)",
+        "in_width": 1024,
+        "in_height": 1024,
+        "in_seed": 0,
+        "in_randomize_seed": true
       }
     }
   ],
   "subjects": [
     {
-      "id": "sub_gallery",
-      "label": "Output Gallery",
+      "id": "sub_image",
+      "label": "Output Image",
+      "role": "subject",
+      "asset_type": "image",
+      "inputs": [{ "id": "in", "label": "Image", "type": "image" }],
+      "outputs": [{ "id": "out", "label": "Image", "type": "image" }],
+      "x": 740,
+      "y": 60,
+      "width": 240,
+      "height": 107,
+      "data": {}
+    },
+    {
+      "id": "sub_seed",
+      "label": "Seed",
       "role": "subject",
-      "asset_type": "gallery",
-      "inputs": [{ "id": "in", "label": "Gallery", "type": "gallery" }],
-      "outputs": [{ "id": "out", "label": "Gallery", "type": "gallery" }],
-      "x": 720,
-      "y": 160,
+      "asset_type": "number",
+      "inputs": [{ "id": "in", "label": "Number", "type": "number" }],
+      "outputs": [{ "id": "out", "label": "Number", "type": "number" }],
+      "x": 740,
+      "y": 200,
       "width": 240,
-      "height": 140,
+      "height": 107,
+      "data": {}
+    },
+    {
+      "id": "sub_caption",
+      "label": "Caption",
+      "role": "subject",
+      "asset_type": "json",
+      "inputs": [{ "id": "in", "label": "JSON", "type": "json" }],
+      "outputs": [{ "id": "out", "label": "JSON", "type": "json" }],
+      "x": 740,
+      "y": 340,
+      "width": 240,
+      "height": 107,
       "data": {}
     }
   ],
   "edges": [
     {
-      "id": "e_prompt_to_ideogram",
+      "id": "e_prompt",
       "from_node_id": "ref_prompt",
       "from_port_id": "out",
-      "to_node_id": "fn_Ideogram 4 (fal)",
+      "to_node_id": "op_ideogram4",
       "to_port_id": "in_prompt",
       "type": "text"
     },
     {
-      "id": "e_ideogram_to_gallery",
-      "from_node_id": "fn_Ideogram 4 (fal)",
-      "from_port_id": "out_0",
-      "to_node_id": "sub_gallery",
+      "id": "e_image",
+      "from_node_id": "op_ideogram4",
+      "from_port_id": "out_image",
+      "to_node_id": "sub_image",
+      "to_port_id": "in",
+      "type": "image"
+    },
+    {
+      "id": "e_seed",
+      "from_node_id": "op_ideogram4",
+      "from_port_id": "out_seed",
+      "to_node_id": "sub_seed",
+      "to_port_id": "in",
+      "type": "number"
+    },
+    {
+      "id": "e_caption",
+      "from_node_id": "op_ideogram4",
+      "from_port_id": "out_caption",
+      "to_node_id": "sub_caption",
       "to_port_id": "in",
-      "type": "gallery"
+      "type": "json"
     }
   ]
 }