From dda6a1db6348384133a9d4fa154384db0991cc8b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 17 Jul 2026 13:08:53 +0000
Subject: [PATCH] Surface HiDream I1 in the image model catalog

One catalog group with the three official bf16 pipelines (Full, plus the Dev
and Fast distillations as labeled artifacts) at their ~63 GB resident size, so
auto-routing keeps this a datacenter-GPU pick. city96's GGUF is deliberately
not wired: the GGUF path would need the same Llama TE4 assembly for very small
demand. Images-page defaults mirror the backend table with the variant keys
ahead of the generic hidream key.
---
 .../model-selector/model-catalog.check.ts     | 16 +++++++++++++
 .../model-selector/model-catalog.ts           | 23 +++++++++++++++++++
 .../src/features/images/images-page.tsx       |  5 ++++
 3 files changed, 44 insertions(+)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.check.ts b/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.check.ts
index 4c98115703..71e6586b7e 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.check.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.check.ts
@@ -399,6 +399,22 @@ assert.equal(
   "bf16",
 );
 assert.equal(groupForRepoId("QuantStack/HunyuanImage-2.1-GGUF", IMAGE_CATALOG), hyimage);
+// HiDream I1: all three variants group together; a datacenter GPU auto-routes to the
+// Full bf16 (catalog order wins among equal sizes), and the group is hidden by the fit
+// filter on a 24 GB card (no GGUF artifact, 63 GB everywhere).
+const hidream = groupForRepoId("HiDream-ai/HiDream-I1-Full", IMAGE_CATALOG);
+assert.ok(hidream);
+assert.equal(groupForRepoId("HiDream-ai/HiDream-I1-Dev", IMAGE_CATALOG), hidream);
+assert.equal(groupForRepoId("HiDream-ai/HiDream-I1-Fast", IMAGE_CATALOG), hidream);
+assert.equal(
+  pickDefaultArtifact(hidream, { gpuGb: 141, systemRamGb: 128, isDownloaded: notDownloaded })
+    .repoId,
+  "HiDream-ai/HiDream-I1-Full",
+);
+assert.equal(
+  catalogGroupFitsDevice(hidream, { gpuGb: 24, systemRamGb: 32 }, notDownloaded),
+  false,
+);
 // FLUX.1-schnell is Apache-2.0 (not gated): its BF16 IS auto-routed on a GPU that fits it.
 const fluxSchnellRoute = groupForRepoId("unsloth/FLUX.1-schnell", IMAGE_CATALOG);
 assert.ok(fluxSchnellRoute);
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.ts b/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.ts
index 5374a07678..0c568bc345 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-catalog.ts
@@ -276,6 +276,29 @@ export const IMAGE_CATALOG: CatalogGroup[] = [
       gguf("QuantStack/HunyuanImage-2.1-GGUF"),
     ],
   },
+  {
+    // 17B MoE DiT + four text encoders. The repos are open (MIT weights) but ship no
+    // Llama text_encoder_4; the backend assembles it from the open unsloth mirror at
+    // load time, adding ~16 GB: ~63 GB bf16-resident total, so this stays a
+    // datacenter-GPU pick. Full is the undistilled base; Dev and Fast are its
+    // guidance-free distillations with their own step defaults. city96's GGUF is not
+    // wired: the GGUF path would need the same TE4 assembly for tiny demand.
+    canonicalId: "HiDream-ai/HiDream-I1-Full",
+    displayName: "HiDream I1",
+    description: "Text-to-image",
+    scope: "image",
+    artifacts: [
+      bf16Pipeline("HiDream-ai/HiDream-I1-Full", 63),
+      bf16Pipeline("HiDream-ai/HiDream-I1-Dev", 63, {
+        label: "BF16 - Dev (distilled)",
+        keywords: ["bf16", "dev", "distilled"],
+      }),
+      bf16Pipeline("HiDream-ai/HiDream-I1-Fast", 63, {
+        label: "BF16 - Fast (distilled)",
+        keywords: ["bf16", "fast", "distilled"],
+      }),
+    ],
+  },
   {
     // No bf16 repo exists for Ideogram 4: -fp8 stores its two DiTs as raw
     // float8 (~46 GB resident after the bf16 cast); -nf4-diffusers is the
diff --git a/studio/frontend/src/features/images/images-page.tsx b/studio/frontend/src/features/images/images-page.tsx
index ee2df72ea7..9ab6c3f74c 100644
--- a/studio/frontend/src/features/images/images-page.tsx
+++ b/studio/frontend/src/features/images/images-page.tsx
@@ -173,6 +173,11 @@ const MODEL_DEFAULTS: Array<{ match: string; steps: number; guidance: number }>
   // HunyuanImage 2.1: 50 steps; the guidance value feeds distilled_guidance_scale
   // (default 3.25), real CFG runs inside the repo's guider components.
   { match: "hunyuanimage", steps: 50, guidance: 3.25 },
+  // HiDream-I1: Full runs 50 steps at guidance 5; the Dev/Fast distillations are
+  // guidance-free. Specific keys before the generic "hidream" (Full + fallback).
+  { match: "hidream-i1-dev", steps: 28, guidance: 0 },
+  { match: "hidream-i1-fast", steps: 16, guidance: 0 },
+  { match: "hidream", steps: 50, guidance: 5 },
   // SDXL: Turbo is distilled (few steps, no CFG); base/full SDXL wants ~30 steps and
   // real CFG (~7). "sdxl-turbo" must precede the generic "sdxl" substring match.
   { match: "sdxl-turbo", steps: 3, guidance: 0 },