Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: generate images through an external connection (#13756)

Nilay2f412e3484a0
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 6d34efb5751..466bc63744e 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -7494,6 +7494,22 @@ class ExternalProviderClient:
raise asyncio.CancelledError
return audio, media_type or f"audio/{response_format}"
+ async def create_images(self, prompt: str, model: str, n: int, size: str) -> list[str]:
+ response = await _client().post(
+ _append_provider_path(self.base_url, "/images/generations"),
+ headers = self._auth_headers(),
+ json = {
+ "model": model,
+ "prompt": prompt,
+ "n": n,
+ "size": size,
+ "response_format": "b64_json",
+ },
+ timeout = self._timeout,
+ )
+ response.raise_for_status()
+ return [item["b64_json"] for item in response.json()["data"]]
+
async def create_transcription(
self,
audio: bytes,
diff --git a/studio/backend/core/inference/image_gallery.py b/studio/backend/core/inference/image_gallery.py
index c2928ca5da7..dfd57d12d6c 100644
--- a/studio/backend/core/inference/image_gallery.py
+++ b/studio/backend/core/inference/image_gallery.py
@@ -51,6 +51,11 @@ def _params_text(meta: dict[str, Any]) -> str:
lines.append(f"Negative prompt: {negative}")
from core.inference.diffusion_sampler import sampler_label, schedule_label
+ if meta.get("model_kind") == "connection":
+ lines.append(
+ f"Size: {meta.get('width')}x{meta.get('height')}, Model: {meta.get('model', '')}"
+ )
+ return "\n".join(lines)
sampler = sampler_label(meta.get("sampler"))
schedule = schedule_label(meta.get("schedule_type"))
lines.append(
diff --git a/studio/backend/core/inference/llama_keepwarm.py b/studio/backend/core/inference/llama_keepwarm.py
index 1c3928af8b9..d2dcf2bd972 100644
--- a/studio/backend/core/inference/llama_keepwarm.py
+++ b/studio/backend/core/inference/llama_keepwarm.py
@@ -403,6 +403,24 @@ def mark_current_response_failed() -> None:
mark_response_failed(_current_response_scope.get())
+_MEDIA_TRACKED_SCOPE_KEY = "_unsloth_keepwarm_media_owner"
+_MEDIA_UNTRACKED_SCOPE_KEY = "_unsloth_keepwarm_media_untracked"
+
+
+def untrack_current_media_request(scope) -> None:
+ """Drop this request from its media tracker once the route knows it won't use the local
+ engine, so a call to a saved connection can't hold off a media swap or idle unload."""
+ if not isinstance(scope, dict) or scope.get(_MEDIA_UNTRACKED_SCOPE_KEY):
+ return
+ owner = scope.get(_MEDIA_TRACKED_SCOPE_KEY)
+ if owner is None:
+ return
+ from core.inference import media_keepwarm
+
+ scope[_MEDIA_UNTRACKED_SCOPE_KEY] = True
+ media_keepwarm.end_request(owner, counted = False)
+
+
def untrack_current_request(scope) -> None:
"""Drop this request from the in-flight count once the route knows it won't use the
local GGUF, so external-provider traffic can't trip the swap busy guard. Idempotent;
@@ -711,6 +729,8 @@ class LlamaKeepWarmMiddleware:
if chat_tracked:
_note_untracked_end(is_preview)
raise
+ if isinstance(scope, dict):
+ scope[_MEDIA_TRACKED_SCOPE_KEY] = media_owner
ended = {"done": False}
status = {"code": None}
# Set once the terminal body frame (more_body False) is sent: only a response that completed cleanly adopts
@@ -725,7 +745,9 @@ class LlamaKeepWarmMiddleware:
return
ended["done"] = True
code = status["code"]
- if media_owner is not None:
+ if media_owner is not None and not (
+ isinstance(scope, dict) and scope.get(_MEDIA_UNTRACKED_SCOPE_KEY)
+ ):
media_keepwarm.end_request(media_owner, counted = code not in (401, 403))
if not chat_tracked:
return
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index e0aaf59818b..65175bd4f2e 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -4968,6 +4968,14 @@ class DiffusionGenerateRequest(BaseModel):
description = "Noise schedule for the sampler. Omitted or 'automatic' uses the sampler's "
"usual schedule (Karras for DPM++ 2M, Exponential for DPM++ 2M SDE, else the model's).",
)
+ provider_id: Optional[str] = Field(
+ None,
+ description = "[x-unsloth] Saved connection ID. When set, the prompt goes to that "
+ "connection's /images/generations as `model` and the results are saved to the gallery.",
+ )
+ model: Optional[str] = Field(
+ None, description = "Model name on the connection named by `provider_id`."
+ )
@field_validator("loras")
@classmethod
@@ -5209,6 +5217,11 @@ class DiffusionGenerateProgressResponse(BaseModel):
None, description = "Live latent preview of the image being denoised, as a JPEG data URL"
)
preview_seq: int = Field(0, description = "Moves each time a new preview is published")
+ target: Optional[Literal["connection"]] = Field(
+ None,
+ description = "'connection' when the run is on a saved image connection, so a page that "
+ "resumed it sends Stop to POST /images/generate/cancel?target=connection",
+ )
class DiffusionLoadProgressResponse(BaseModel):
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 15ccc8a2d10..2af2bfddec6 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -47613,11 +47613,209 @@ def _generate_failure_detail(message: str) -> str:
return _GENERATE_FAILURE_FALLBACK
+# Keyed by account id, not account_scope(): the scope moves when the first managed account appears.
+_external_image_runs: dict[str, set] = {}
+_stopped_external_image_runs: set = set()
+
+
+def _external_image_run_active() -> bool:
+ return bool(_external_image_runs.get(current_account_id()))
+
+
+def _stop_external_image_runs() -> bool:
+ runs = [
+ task for task in _external_image_runs.get(current_account_id()) or () if not task.done()
+ ]
+ for task in runs:
+ _stopped_external_image_runs.add(task)
+ task.cancel()
+ return bool(runs)
+
+
+async def _external_image_generation(
+ body: DiffusionGenerateRequest, request: Request
+) -> DiffusionGenerateResponse:
+ from PIL import Image
+
+ from core.inference import image_gallery
+ from core.inference.llama_keepwarm import (
+ untrack_current_media_request,
+ untrack_current_request,
+ )
+
+ untrack_current_request(request.scope)
+ untrack_current_media_request(request.scope)
+ provider_id = body.provider_id.strip()
+ external_model = (body.model or "").strip()
+ if not external_model:
+ raise HTTPException(
+ status_code = 400,
+ detail = "model is required when using an external image connection.",
+ )
+ if (
+ body.init_image
+ or body.reference_images
+ or body.loras
+ or body.controlnet
+ or body.prompts
+ or body.seeds
+ or body.negative_prompt
+ or body.mask_image
+ or body.strength is not None
+ or body.upscale is not None
+ or body.workflow
+ or body.localized_edit
+ or body.sampler
+ or body.schedule_type
+ ):
+ raise HTTPException(
+ status_code = 400,
+ detail = "An external image connection only supports text to image.",
+ )
+
+ config = await asyncio.to_thread(providers_db.get_provider, provider_id)
+ if config is None:
+ raise HTTPException(status_code = 404, detail = f"Provider config not found: {provider_id}")
+ _refuse_decision_connection(config["provider_type"], config.get("api_type"))
+ if not config["is_enabled"]:
+ raise HTTPException(
+ status_code = 400,
+ detail = f"Provider '{config['display_name']}' is disabled.",
+ )
+ base_url = config["base_url"] or get_base_url(config["provider_type"])
+ if not base_url:
+ raise HTTPException(status_code = 400, detail = "The connection has no base URL.")
+ try:
+ base_url = validate_provider_base_url(base_url)
+ except ValueError as exc:
+ raise HTTPException(status_code = 400, detail = str(exc)) from None
+
+ async with provider_config_guard(provider_id):
+ current = await asyncio.to_thread(providers_db.get_provider, provider_id)
+ routing_fields = ("provider_type", "base_url", "is_enabled")
+ if current is None or any(
+ current.get(field) != config.get(field) for field in routing_fields
+ ):
+ raise HTTPException(
+ status_code = 409,
+ detail = "Provider changed while the request was starting; retry.",
+ )
+ api_key = await asyncio.to_thread(
+ resolve_provider_api_key_or_400,
+ provider_id,
+ None,
+ allow_saved_key = not _request_has_api_key(request),
+ )
+ latest = await asyncio.to_thread(providers_db.get_provider, provider_id)
+ if latest is None or any(
+ latest.get(field) != current.get(field) for field in routing_fields
+ ):
+ raise HTTPException(
+ status_code = 409,
+ detail = "Provider changed while the request was starting; retry.",
+ )
+
+ client = ExternalProviderClient(
+ provider_type = config["provider_type"],
+ base_url = base_url,
+ api_key = api_key,
+ timeout = 600.0,
+ )
+ owner = current_account_id()
+ task = asyncio.create_task(
+ client.create_images(
+ prompt = body.prompt,
+ model = external_model,
+ n = body.batch_size,
+ size = f"{body.width}x{body.height}",
+ )
+ )
+ _external_image_runs.setdefault(owner, set()).add(task)
+ try:
+ try:
+ try:
+ encoded = await task
+ except asyncio.CancelledError:
+ if task not in _stopped_external_image_runs:
+ task.cancel()
+ raise
+ from core.inference.diffusion_families import DIFFUSION_CANCELLED_MSG
+ raise HTTPException(status_code = 409, detail = DIFFUSION_CANCELLED_MSG) from None
+ images = []
+ for item in encoded:
+ image = Image.open(io.BytesIO(base64.b64decode(item)))
+ image.load()
+ images.append(image)
+ except HTTPException:
+ raise
+ except httpx.HTTPStatusError as exc:
+ raise HTTPException(
+ status_code = 502,
+ detail = (
+ f"Image endpoint returned HTTP {exc.response.status_code}: {exc.response.text[:300]}"
+ ),
+ ) from exc
+ except httpx.HTTPError as exc:
+ raise log_and_http_error(
+ exc,
+ 502,
+ "Could not reach the image endpoint.",
+ event = "external_images.transport_error",
+ log = logger,
+ ) from exc
+ except Exception as exc:
+ raise HTTPException(
+ status_code = 502, detail = "The image endpoint did not return a usable image."
+ ) from exc
+
+ created_at = time.time()
+
+ def _persist() -> list[dict]:
+ return [
+ image_gallery.save(
+ image,
+ {
+ "prompt": body.prompt,
+ "width": image.width,
+ "height": image.height,
+ "steps": body.steps,
+ "guidance": body.guidance,
+ "seed": body.seed or 0,
+ "batch_index": index,
+ "batch_size": body.batch_size,
+ "model": external_model,
+ "model_kind": "connection",
+ "created_at": created_at,
+ },
+ )
+ for index, image in enumerate(images)
+ ]
+
+ try:
+ records = await asyncio.to_thread(_persist)
+ except Exception as exc:
+ logger.error("diffusion.persist_failed: %s", exc)
+ raise HTTPException(status_code = 500, detail = "Failed to save the generated image.")
+ return DiffusionGenerateResponse(images = [GalleryImage(**r) for r in records])
+ finally:
+ runs = _external_image_runs.get(owner)
+ if runs is not None:
+ runs.discard(task)
+ if not runs:
+ _external_image_runs.pop(owner, None)
+ _stopped_external_image_runs.discard(task)
+
+
@studio_router.post("/images/generate", response_model = DiffusionGenerateResponse)
@account_access.gpu_busy_route
async def generate_diffusion_image(
- request: DiffusionGenerateRequest, current_subject: str = Depends(get_current_subject)
+ request: DiffusionGenerateRequest,
+ http_request: Request,
+ current_subject: str = Depends(get_current_subject),
):
+ if request.provider_id:
+ return await _external_image_generation(request, http_request)
+
from core.inference import image_gallery
from core.inference.diffusion_engine_router import get_active_diffusion_engine
from core.inference.diffusion_families import (
@@ -48587,11 +48785,18 @@ async def diffusion_load_progress(
@studio_router.get("/images/generate-progress", response_model = DiffusionGenerateProgressResponse)
async def diffusion_generate_progress(current_subject: str = Depends(get_current_subject)):
- if account_access.managed_account() and account_access.generation_is_foreign("diffusion"):
+ external = _external_image_run_active()
+
+ def _hidden():
+ if external:
+ return DiffusionGenerateProgressResponse(active = True, target = "connection")
return account_access.hidden_generate_progress_response(DiffusionGenerateProgressResponse)
+
+ if account_access.managed_account() and account_access.generation_is_foreign("diffusion"):
+ return _hidden()
mine = account_access.generation_is_mine("diffusion")
if not mine and account_access.resident_hidden("diffusion"):
- return account_access.hidden_generate_progress_response(DiffusionGenerateProgressResponse)
+ return _hidden()
from core.inference.diffusion_engine_router import get_active_diffusion_engine
if (
@@ -48601,10 +48806,12 @@ async def diffusion_generate_progress(current_subject: str = Depends(get_current
"diffusion", get_active_diffusion_engine().status().get("repo_id")
)
):
- return account_access.hidden_generate_progress_response(DiffusionGenerateProgressResponse)
+ return _hidden()
progress = get_active_diffusion_engine().generate_progress()
log_media_generation_progress("image", progress)
+ if external and not progress["active"]:
+ progress = {**progress, "active": True, "target": "connection"}
# A finished generation still persisting its gallery record counts as active, so a reload probe keeps polling.
if _diffusion_persist_active > 0 and not progress["active"]:
progress = {**progress, "active": True}
@@ -48612,7 +48819,9 @@ async def diffusion_generate_progress(current_subject: str = Depends(get_current
@studio_router.post("/images/generate/cancel")
-async def cancel_diffusion_generation(current_subject: str = Depends(get_current_subject)):
+async def cancel_diffusion_generation(
+ target: Optional[str] = None, current_subject: str = Depends(get_current_subject)
+):
"""Stop the in-flight image generation, mirroring POST /video/generate/cancel.
Resolved through the engine router, so it stops a diffusers denoise (at its next step
@@ -48620,7 +48829,10 @@ async def cancel_diffusion_generation(current_subject: str = Depends(get_current
generation's OWN request is what reports the outcome: it unwinds with the cancelled
sentinel, which this module already maps to a 409. ``cancelled`` is False when nothing was
running, so the page can settle its button back to Generate rather than wait for a
- generation that already finished."""
+ generation that already finished. ``target=connection`` stops this account's runs on a
+ saved image connection instead."""
+ if target == "connection":
+ return {"cancelled": _stop_external_image_runs()}
if account_access.generation_is_foreign("diffusion"):
return {"cancelled": False}
mine = account_access.generation_is_mine("diffusion")
diff --git a/studio/backend/tests/test_diffusion_routes.py b/studio/backend/tests/test_diffusion_routes.py
index 0da728385ec..baacf0b21c9 100644
--- a/studio/backend/tests/test_diffusion_routes.py
+++ b/studio/backend/tests/test_diffusion_routes.py
@@ -3048,3 +3048,353 @@ def test_a_managed_account_cannot_send_allow_oversized(client, monkeypatch):
)
assert resp.status_code == 200
assert seen == [False]
+
+
+def _external_image_connection(monkeypatch, handler):
+ import httpx
+
+ import core.inference.external_provider as external_provider
+
+ monkeypatch.setattr(
+ external_provider,
+ "_http_client",
+ httpx.AsyncClient(transport = httpx.MockTransport(handler)),
+ )
+ monkeypatch.setattr(
+ inference_routes.providers_db,
+ "get_provider",
+ lambda pid: (
+ {
+ "provider_type": "custom",
+ "display_name": "Gufo",
+ "base_url": "http://127.0.0.1:8000/v1",
+ "is_enabled": True,
+ }
+ if pid == "conn-1"
+ else None
+ ),
+ )
+ monkeypatch.setattr(
+ inference_routes, "resolve_provider_api_key_or_400", lambda *a, **k: "sk-test"
+ )
+
+
+def test_external_connection_generates_without_a_local_model_and_saves_to_gallery(
+ client, monkeypatch
+):
+ import base64
+ import io
+ import json
+
+ import httpx
+ from PIL import Image
+
+ sent = []
+
+ def handler(req):
+ sent.append((str(req.url), req.headers.get("authorization"), json.loads(req.content)))
+ buf = io.BytesIO()
+ Image.new("RGB", (512, 256), (200, 30, 30)).save(buf, format = "PNG")
+ png = base64.b64encode(buf.getvalue()).decode()
+ return httpx.Response(200, json = {"created": 0, "data": [{"b64_json": png}] * 2})
+
+ _external_image_connection(monkeypatch, handler)
+ resp = _post_generate(
+ client,
+ prompt = "a sloth",
+ width = 512,
+ height = 256,
+ batch_size = 2,
+ provider_id = "conn-1",
+ model = "qwen-image",
+ )
+ assert resp.status_code == 200, resp.text
+ images = resp.json()["images"]
+ assert [(i["prompt"], i["model"], i["width"], i["height"]) for i in images] == [
+ ("a sloth", "qwen-image", 512, 256)
+ ] * 2
+ assert len(client.get("/api/inference/images/gallery").json()["images"]) == 2
+ assert sent == [
+ (
+ "http://127.0.0.1:8000/v1/images/generations",
+ "Bearer sk-test",
+ {
+ "model": "qwen-image",
+ "prompt": "a sloth",
+ "n": 2,
+ "size": "512x256",
+ "response_format": "b64_json",
+ },
+ )
+ ]
+
+
+def test_external_connection_error_is_a_502_with_the_upstream_reason(client, monkeypatch):
+ import httpx
+
+ _external_image_connection(
+ monkeypatch, lambda req: httpx.Response(404, text = "model 'nope' not found")
+ )
+ resp = _post_generate(client, prompt = "a sloth", provider_id = "conn-1", model = "nope")
+ assert resp.status_code == 502
+ assert "model 'nope' not found" in resp.text
+ assert client.get("/api/inference/images/gallery").json()["images"] == []
+
+
+def test_external_connection_does_not_hold_local_chat_swaps(client, monkeypatch):
+ import base64
+ import io
+
+ import httpx
+ from PIL import Image
+
+ import core.inference.llama_keepwarm as kw
+
+ client.app.add_middleware(kw.LlamaKeepWarmMiddleware)
+ monkeypatch.setattr(kw, "_inflight", 0)
+ seen = []
+
+ def handler(req):
+ seen.append(
+ kw.other_inference_request_count(current_request_counted = False, include_pending = False)
+ )
+ buf = io.BytesIO()
+ Image.new("RGB", (64, 64)).save(buf, format = "PNG")
+ return httpx.Response(
+ 200, json = {"data": [{"b64_json": base64.b64encode(buf.getvalue()).decode()}]}
+ )
+
+ _external_image_connection(monkeypatch, handler)
+ resp = client.post(
+ "/api/inference/images/generate",
+ json = {"prompt": "a sloth", "provider_id": "conn-1", "model": "qwen-image"},
+ headers = {"Authorization": "Bearer session"},
+ )
+ assert resp.status_code == 200, resp.text
+ assert seen == [0]
+ assert kw._inflight == 0
+
+
+def test_external_connection_refuses_a_key_read_after_the_connection_moved(client, monkeypatch):
+ import httpx
+
+ sent = []
+ _external_image_connection(monkeypatch, lambda req: sent.append(req) or httpx.Response(500))
+ rows = iter(["http://127.0.0.1:8000/v1"] * 2 + ["http://10.0.0.9:9000/v1"])
+ monkeypatch.setattr(
+ inference_routes.providers_db,
+ "get_provider",
+ lambda pid: {
+ "provider_type": "custom",
+ "display_name": "Gufo",
+ "base_url": next(rows),
+ "is_enabled": True,
+ },
+ )
+ resp = _post_generate(client, prompt = "a sloth", provider_id = "conn-1", model = "qwen-image")
+ assert resp.status_code == 409
+ assert "Provider changed" in resp.json()["detail"]
+ assert sent == []
+
+
+def test_stop_cancels_a_connection_run_and_progress_shows_it(client, monkeypatch):
+ import asyncio
+
+ import httpx
+
+ started = asyncio.Event()
+ calls = []
+
+ async def handler(req):
+ calls.append(req)
+ started.set()
+ await asyncio.sleep(30)
+ return httpx.Response(500)
+
+ _external_image_connection(monkeypatch, handler)
+
+ async def run():
+ transport = httpx.ASGITransport(app = client.app)
+ async with httpx.AsyncClient(transport = transport, base_url = "http://t") as ac:
+ gen = asyncio.create_task(
+ ac.post(
+ "/api/inference/images/generate",
+ json = {"prompt": "a sloth", "provider_id": "conn-1", "model": "qwen-image"},
+ )
+ )
+ await asyncio.wait_for(started.wait(), 5)
+ progress = (await ac.get("/api/inference/images/generate-progress")).json()
+ local = (await ac.post("/api/inference/images/generate/cancel")).json()
+ stop = (await ac.post("/api/inference/images/generate/cancel?target=connection")).json()
+ resp = await asyncio.wait_for(gen, 5)
+ after = (await ac.get("/api/inference/images/generate-progress")).json()
+ return progress, local, stop, resp, after
+
+ progress, local, stop, resp, after = asyncio.run(run())
+ assert progress["active"] is True
+ # A page that resumed this run reads its Stop target from progress.
+ assert progress["target"] == "connection"
+ assert local == {"cancelled": False}
+ assert stop == {"cancelled": True}
+ assert resp.status_code == 409
+ assert after["active"] is False
+ assert after["target"] is None
+ assert len(calls) == 1
+ assert client.get("/api/inference/images/gallery").json()["images"] == []
+
+
+def test_stop_reaches_a_connection_run_after_the_account_scope_moves(client, monkeypatch):
+ import asyncio
+
+ import httpx
+ import hub.services.models.account_access as account_access
+
+ started = asyncio.Event()
+
+ async def handler(req):
+ started.set()
+ await asyncio.sleep(30)
+ return httpx.Response(500)
+
+ _external_image_connection(monkeypatch, handler)
+
+ async def run():
+ transport = httpx.ASGITransport(app = client.app)
+ async with httpx.AsyncClient(transport = transport, base_url = "http://t") as ac:
+ gen = asyncio.create_task(
+ ac.post(
+ "/api/inference/images/generate",
+ json = {"prompt": "a sloth", "provider_id": "conn-1", "model": "qwen-image"},
+ )
+ )
+ await asyncio.wait_for(started.wait(), 5)
+ # The first managed account appears mid-run: the owner's scope goes from None to its id.
+ monkeypatch.setattr(account_access, "account_scope", lambda: "owner-scope")
+ progress = (await ac.get("/api/inference/images/generate-progress")).json()
+ stop = (await ac.post("/api/inference/images/generate/cancel?target=connection")).json()
+ resp = await asyncio.wait_for(gen, 5)
+ return progress, stop, resp
+
+ progress, stop, resp = asyncio.run(run())
+ assert progress["active"] is True
+ assert stop == {"cancelled": True}
+ assert resp.status_code == 409
+
+
+def test_external_connection_refuses_options_it_cannot_forward(client, monkeypatch):
+ import httpx
+
+ sent = []
+ _external_image_connection(monkeypatch, lambda req: sent.append(req) or httpx.Response(500))
+ for extra in ({"negative_prompt": "blurry"}, {"sampler": "euler"}, {"strength": 0.5}):
+ resp = _post_generate(
+ client, prompt = "a sloth", provider_id = "conn-1", model = "qwen-image", **extra
+ )
+ assert resp.status_code == 400, (extra, resp.text)
+ assert sent == []
+
+
+def test_connection_images_are_marked_in_the_gallery(client, monkeypatch):
+ import base64
+ import io
+
+ import httpx
+ from PIL import Image
+
+ def handler(req):
+ buf = io.BytesIO()
+ Image.new("RGB", (64, 64)).save(buf, format = "PNG")
+ return httpx.Response(
+ 200, json = {"data": [{"b64_json": base64.b64encode(buf.getvalue()).decode()}]}
+ )
+
+ _external_image_connection(monkeypatch, handler)
+ resp = _post_generate(client, prompt = "a sloth", provider_id = "conn-1", model = "qwen-image")
+ assert resp.status_code == 200, resp.text
+ assert resp.json()["images"][0]["model_kind"] == "connection"
+
+
+def test_a_connection_run_reports_active_even_when_local_progress_is_hidden(client, monkeypatch):
+ import hub.services.models.account_access as account_access
+
+ monkeypatch.setattr(account_access, "managed_account", lambda: True)
+ monkeypatch.setattr(account_access, "generation_is_foreign", lambda kind: True)
+ assert client.get("/api/inference/images/generate-progress").json()["active"] is False
+ from utils.account_context import current_account_id
+
+ monkeypatch.setitem(inference_routes._external_image_runs, current_account_id(), {object()})
+ progress = client.get("/api/inference/images/generate-progress").json()
+ assert progress["active"] is True
+ assert progress["target"] == "connection"
+
+
+def test_a_connection_call_leaves_the_media_tracker(client, monkeypatch):
+ import base64
+ import io
+
+ import httpx
+ from PIL import Image
+
+ import core.inference.llama_keepwarm as kw
+ import core.inference.media_keepwarm as mk
+
+ client.app.add_middleware(kw.LlamaKeepWarmMiddleware)
+ seen = []
+
+ def handler(req):
+ seen.append(mk._TRACKERS[mk.DIFFUSION].outstanding())
+ buf = io.BytesIO()
+ Image.new("RGB", (64, 64)).save(buf, format = "PNG")
+ return httpx.Response(
+ 200, json = {"data": [{"b64_json": base64.b64encode(buf.getvalue()).decode()}]}
+ )
+
+ _external_image_connection(monkeypatch, handler)
+ before = mk._TRACKERS[mk.DIFFUSION].outstanding()
+ resp = client.post(
+ "/api/inference/images/generate",
+ json = {"prompt": "a sloth", "provider_id": "conn-1", "model": "qwen-image"},
+ headers = {"Authorization": "Bearer session"},
+ )
+ assert resp.status_code == 200, resp.text
+ assert seen == [before]
+ assert mk._TRACKERS[mk.DIFFUSION].outstanding() == before
+
+
+def test_stop_after_the_connection_answered_is_not_acknowledged(client, monkeypatch):
+ import asyncio
+
+ from utils.account_context import current_account_id
+
+ async def run():
+ done = asyncio.get_running_loop().create_future()
+ done.set_result(None)
+ task = asyncio.ensure_future(done)
+ await task
+ inference_routes._external_image_runs[current_account_id()] = {task}
+ try:
+ return inference_routes._stop_external_image_runs()
+ finally:
+ inference_routes._external_image_runs.clear()
+
+ assert asyncio.run(run()) is False
+
+
+def test_connection_png_parameters_leave_out_unsent_settings():
+ from core.inference.image_gallery import _params_text
+
+ text = _params_text(
+ {
+ "prompt": "a sloth",
+ "steps": 9,
+ "guidance": 0.0,
+ "seed": 123,
+ "width": 512,
+ "height": 256,
+ "model": "qwen-image",
+ "model_kind": "connection",
+ }
+ )
+ assert text == "a sloth\nSize: 512x256, Model: qwen-image"
+ local = _params_text({"prompt": "a sloth", "steps": 9, "guidance": 0.0, "seed": 123})
+ assert "Steps: 9" in local and "Seed: 123" in local
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index e6088b433bc..75ccf8fe9f5 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -328,6 +328,7 @@ export { ModelLoadDescription } from "./components/model-load-status";
export { chatModelLoaded } from "./lib/chat-model-loaded";
export type { ChatModelLoadedInput } from "./lib/chat-model-loaded";
export {
+ buildExternalModelId,
customProviderDisplayName,
isCustomProviderType,
isDecisionConnection,
diff --git a/studio/frontend/src/features/images/api.ts b/studio/frontend/src/features/images/api.ts
index bc08b8c2ad2..a2b1fecf9a8 100644
--- a/studio/frontend/src/features/images/api.ts
+++ b/studio/frontend/src/features/images/api.ts
@@ -116,6 +116,8 @@ export interface DiffusionGenerateProgress {
// Live latent preview of the image being denoised (small JPEG data URL), and a counter that moves with each one.
preview?: string | null;
preview_seq?: number;
+ // "connection" while the run is on a saved image connection: the Stop target for a resumed run.
+ target?: "connection" | null;
}
export interface DiffusionLoadProgress {
@@ -449,10 +451,7 @@ export async function generateDiffusionImage(
// A proxy answers with HTML (or nothing); the app answers with JSON. Only the former means the request may still
// be running: settling an application error would poll for a generation that never started and hide the reason
// the backend gave.
- if (
- response.status === 503 &&
- (response.headers.get("content-type") || "").toLowerCase().includes("application/json")
- ) {
+ if ((response.headers.get("content-type") || "").toLowerCase().includes("application/json")) {
throw new Error(detail);
}
throw new GenerateResponseLostError(detail);
diff --git a/studio/frontend/src/features/images/connection-cancel.ts b/studio/frontend/src/features/images/connection-cancel.ts
new file mode 100644
index 00000000000..0112eb4c6a6
--- /dev/null
+++ b/studio/frontend/src/features/images/connection-cancel.ts
@@ -0,0 +1,19 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+/** Stop this account's runs on a saved image connection. Kept out of api.ts, which is in the startup
+ * bundle; same no-retry, abortable contract as cancelDiffusionGeneration. */
+export async function cancelConnectionImageGeneration(
+ signal?: AbortSignal,
+): Promise<{ cancelled: boolean }> {
+ const response = await authFetch(
+ "/api/inference/images/generate/cancel?target=connection",
+ { method: "POST", signal },
+ { retryNetworkErrors: false },
+ );
+ if (!response.ok) throw new Error(await readFastApiError(response));
+ return (await response.json()) as { cancelled: boolean };
+}
diff --git a/studio/frontend/src/features/images/images-page.tsx b/studio/frontend/src/features/images/images-page.tsx
index 9f8510539c3..3eace9bb9ec 100644
--- a/studio/frontend/src/features/images/images-page.tsx
+++ b/studio/frontend/src/features/images/images-page.tsx
@@ -96,6 +96,7 @@ import {
} from "@/hooks/use-host-class";
import { nvfp4SelectionFallback, withNvfp4Option } from "@/lib/nvfp4-options";
import type {
+ ExternalModelOption,
ModelOption,
ModelSelectorChangeMeta,
} from "@/features/model-picker/components/model-selector/types";
@@ -145,8 +146,15 @@ import {
recipeWorkflowLabel,
type WorkflowId,
} from "./workflows";
-import { ParamSlider } from "@/features/chat";
+import {
+ ParamSlider,
+ buildExternalModelId,
+ isDecisionConnection,
+ parseExternalModelId,
+ useExternalProvidersStore,
+} from "@/features/chat";
import { ModelLoadDescription } from "@/features/chat/components/model-load-status";
+import { cancelConnectionImageGeneration } from "./connection-cancel";
import {
type ImageGenerationPresetParams,
MediaGenerationPresetControl,
@@ -720,6 +728,7 @@ function loadToastArgs(
}
const NO_SAMPLERS: DiffusionSamplerOption[] = [];
+const IMAGE_CONNECTION_TYPES = new Set(["custom", "vllm", "llama_cpp", "lemonade", "ollama"]);
const IDLE_PROGRESS: DiffusionLoadProgress = {
phase: null,
@@ -1194,6 +1203,7 @@ function RecipePopover({
useEffect(() => {
if (!active) setOpen(false);
}, [active]);
+ const viaConnection = image.model_kind === "connection";
return (
<Popover open={active && open} onOpenChange={(o) => setOpen(active && o)}>
<PopoverTrigger asChild>
@@ -1290,15 +1300,15 @@ function RecipePopover({
/>
) : null}
<RecipeRow label="Size" value={`${image.width} × ${image.height}`} />
- <RecipeRow label="Steps" value={String(image.steps)} />
- <RecipeRow label="Guidance" value={String(image.guidance)} />
+ {!viaConnection && <RecipeRow label="Steps" value={String(image.steps)} />}
+ {!viaConnection && <RecipeRow label="Guidance" value={String(image.guidance)} />}
{samplerRecipeLabel(image.sampler, image.schedule_type) ? (
<RecipeRow
label="Sampler"
value={samplerRecipeLabel(image.sampler, image.schedule_type) ?? ""}
/>
) : null}
- <RecipeRow label="Seed" value={String(image.seed)} mono />
+ {!viaConnection && <RecipeRow label="Seed" value={String(image.seed)} mono />}
</div>
<div className="shrink-0 border-t border-border/60 px-3 py-2.5">
<Button size="sm" className="w-full gap-1.5" onClick={() => onRestore(image)}>
@@ -1459,6 +1469,34 @@ export function ImagesPage({
const nvfp4Diffusion = useNvfp4Diffusion();
const nvfp4DiffusionKnown = useNvfp4DiffusionKnown();
const imageModels = useImageModels(hostClass, denseQuantSchemes);
+ const connections = useExternalProvidersStore((s) => s.providers);
+ const connectionsEnabled = useExternalProvidersStore((s) => s.connectionsEnabled);
+ const externalModels = useMemo<ExternalModelOption[]>(
+ () =>
+ connectionsEnabled
+ ? connections
+ .filter((c) => IMAGE_CONNECTION_TYPES.has(c.providerType) && !isDecisionConnection(c))
+ .flatMap((c) =>
+ c.models.map((model) => ({
+ id: buildExternalModelId(c.id, model),
+ name: model,
+ providerId: c.id,
+ providerName: c.name,
+ providerType: c.providerType,
+ })),
+ )
+ : [],
+ [connections, connectionsEnabled],
+ );
+ const [externalModelId, setExternalModelId] = useState<string | null>(null);
+ const runTarget = useRef<"connection" | undefined>(undefined);
+ const external = useMemo(
+ () =>
+ externalModels.some((m) => m.id === externalModelId)
+ ? parseExternalModelId(externalModelId)
+ : null,
+ [externalModels, externalModelId],
+ );
const { rootStyle: railRootStyle } = useMediaRailWidth("images");
const [quant, setQuant] = useState<string | null>(galleryCache.quant);
// One prompt per workflow, starting from the last one generated with.
@@ -1668,7 +1706,8 @@ export function ImagesPage({
const genVisibilityListener = useRef<(() => void) | null>(null);
const [status, setStatus] = useState<DiffusionStatus | null>(null);
const { familyOverride, setFamilyOverride, familySelect, opaqueKind, selectorModelId } = useFamilyOverride(status, status?.supported_families);
- const conditioning = status?.loaded ? (status.conditioning ?? null) : null;
+ // A connection is not bound by the resident model's size grid or presets.
+ const conditioning = !external && status?.loaded ? (status.conditioning ?? null) : null;
const sizeLimits = useMemo(() => sizeLimitsFrom(conditioning), [conditioning]);
const unifiedEdit = Boolean(conditioning?.unified_edit);
const referenceResolutions = conditioning?.reference_resolutions ?? [];
@@ -1896,7 +1935,7 @@ export function ImagesPage({
// Refresh the LoRA picker when the loaded family changes, since a LoRA is family-specific.
// Not on first load or unload (a restore can precede the load), so track it in a ref.
- const loraCapable = Boolean(status?.loaded && status?.supports_lora);
+ const loraCapable = Boolean(!external && status?.loaded && status?.supports_lora);
const prevLoraFamilyRef = useRef<string | null | undefined>(undefined);
// Whether the load in flight baked the LoRA selection into the build (see handleLoad).
const bakedLorasOnLoad = useRef(false);
@@ -1960,9 +1999,10 @@ export function ImagesPage({
}, [loraCapable, residentBuildKey, status?.resolved, loras]);
// Refresh the ControlNet options when the loaded family changes, and clear a stale selection.
- const controlnetCapable = Boolean(status?.loaded && status?.supports_controlnet);
- const negativeCapable = status?.supports_negative_prompt !== false;
- const samplerOptions = (status?.loaded ? status.samplers : null) ?? NO_SAMPLERS;
+ const controlnetSupported = Boolean(status?.loaded && status?.supports_controlnet);
+ const controlnetCapable = !external && controlnetSupported;
+ const negativeCapable = !external && status?.supports_negative_prompt !== false;
+ const samplerOptions = (!external && status?.loaded ? status.samplers : null) ?? NO_SAMPLERS;
// The checkpoint, not residentDefaults: community SDXL files all share the SDXL base repo.
const samplerModelKey = status?.loaded ? `${status.repo_id ?? ""}|${status.gguf_filename ?? ""}` : "";
const samplerChoice = normalizeSamplerChoice(
@@ -1983,7 +2023,7 @@ export function ImagesPage({
[samplerChoice, samplerModelKey, samplerOptions],
);
useEffect(() => {
- if (!controlnetCapable) {
+ if (!controlnetSupported) {
setAvailableControlNets([]);
setControlnetId("");
setControlImage(null);
@@ -2002,7 +2042,7 @@ export function ImagesPage({
return () => {
cancelled = true;
};
- }, [controlnetCapable, status?.family]);
+ }, [controlnetSupported, status?.family]);
// The control types offered for the selected ControlNet: a union model advertises several.
const controlTypeOptions = useMemo(() => {
@@ -2493,12 +2533,15 @@ export function ImagesPage({
);
const restoreSettings = useCallback((image: GalleryImage) => {
+ const viaConnection = image.model_kind === "connection";
// Negative prompt only applies when guidance>0; do not restore a hidden value.
const restoredNegative = image.guidance > 0 ? (image.negative_prompt ?? "") : "";
setNegativePrompt(restoredNegative);
if (restoredNegative) setNegativeOpen(true);
- setSteps(image.steps);
- setGuidance(image.guidance);
+ if (!viaConnection) {
+ setSteps(image.steps);
+ setGuidance(image.guidance);
+ }
const restoredChoice = {
sampler: image.sampler ?? DEFAULT_SAMPLER,
scheduleType: image.schedule_type ?? AUTOMATIC_SCHEDULE,
@@ -2507,7 +2550,7 @@ export function ImagesPage({
else setRestoredSampler(restoredChoice);
// Restore from the BASE batch seed, not this image's derived seed, or a replay with batch_size
// advances it again.
- setSeed(String(image.batch_seed ?? image.seed));
+ if (!viaConnection) setSeed(String(image.batch_seed ?? image.seed));
const restored = restorableSize(image.width, image.height, image.workflow, sizeLimits);
setWidth(restored.width);
setHeight(restored.height);
@@ -2807,6 +2850,7 @@ export function ImagesPage({
pollInFlight = true;
try {
const p = await getGenerateProgress();
+ if (p.active) runTarget.current = p.target ?? undefined;
if (!p.active) {
if (genPollTimer.current) clearInterval(genPollTimer.current);
genPollTimer.current = null;
@@ -2859,6 +2903,7 @@ export function ImagesPage({
try {
const g = await getGenerateProgress();
if (g.active) {
+ runTarget.current = g.target ?? undefined;
setBusy("generating");
setGenStep(g);
resumeGeneratePoll();
@@ -3039,6 +3084,7 @@ export function ImagesPage({
// Reuse the pick toast when loading starts.
pickToastId?: string,
): Promise<boolean> => {
+ setExternalModelId(null);
// Cancel any prior poll loop so two cannot run at once.
if (pollTimer.current) clearTimeout(pollTimer.current);
// Read BEFORE the start request goes out: a Cancel pressed while it is in flight sends an
@@ -3582,7 +3628,10 @@ export function ImagesPage({
// already staging with its claim, or resumePendingLoad drops that load once its files arrive.
const downloadOnlyPick = modelSelectionAction === "download";
const token = downloadOnlyPick ? undefined : pickGuard.claim();
- if (!downloadOnlyPick) setFamilyOverride("auto");
+ if (!downloadOnlyPick) {
+ setFamilyOverride("auto");
+ setExternalModelId(null);
+ }
void navigateSelf({ to: "/images", search: {}, replace: true });
// A label means a GGUF repo whatever the catalog says, and is not loadable, so resolve it
// rather than routing it as a filename.
@@ -3696,6 +3745,19 @@ export function ImagesPage({
// with a 409. A download-only pick submits no load, so that cannot happen, and the selector
// stays interactive during a generation: refusing it there was a silent dead click.
if (busy !== null && !downloadOnlyPick) return;
+ if (meta.source === "external") {
+ if (busy !== null) return;
+ if (quantRevert.current && quantRevert.current === stagedQuantRevert.current) {
+ revertPick(quantRevert.current);
+ quantRevert.current = null;
+ }
+ pickGuard.cancel();
+ pickToast.dismissAll();
+ beginPick();
+ setExternalModelId(id);
+ return;
+ }
+ if (!downloadOnlyPick) setExternalModelId(null);
if (!downloadOnlyPick) beginPick();
// This pick owns the page now, so one still awaiting a listing or a plan drops out. Before any
// branch, since staging never sets `busy`.
@@ -3883,6 +3945,7 @@ export function ImagesPage({
loadGgufRepoPick,
loadOrStage,
pickGuard,
+ pickToast,
quant,
revertPick,
],
@@ -3906,6 +3969,7 @@ export function ImagesPage({
// The deploy owns the page now: a resolving pick or a staged download would load over the base it is about to.
pickGuard.cancel();
pickToast.dismissAll();
+ setExternalModelId(null);
pendingDeploy.current = { loraId: stem, family: args.family };
if (args.trigger.trim()) setPrompt(args.trigger.trim());
setPageMode("create");
@@ -4375,7 +4439,7 @@ export function ImagesPage({
}
// Resolve a base seed up front, so each sequential image gets a distinct, reproducible seed (base + i).
let baseSeed: number;
- if (seed.trim()) {
+ if (seed.trim() && !external) {
const n = Number(seed);
if (!Number.isInteger(n) || n < 0 || n > Number.MAX_SAFE_INTEGER) {
toast.error("Seed must be a non-negative integer");
@@ -4422,6 +4486,7 @@ export function ImagesPage({
cancelAbort.current?.abort();
cancelAbort.current = null;
runToken.current += 1;
+ runTarget.current = external ? "connection" : undefined;
// Poll the backend's per-step progress across the whole run. A named poll body also serves the
// visibilitychange listener, so a throttled tab catches up when visible.
let pollInFlight = false;
@@ -4515,6 +4580,7 @@ export function ImagesPage({
strength: controlStrength,
}
: undefined,
+ ...(external && { provider_id: external.providerId, model: external.modelId }),
});
} catch (err) {
// The POST response was lost while the backend kept generating (the secure-mode tunnel caps
@@ -4581,7 +4647,7 @@ export function ImagesPage({
setGenStep(null);
setStopping(false);
}
- }, [allowOversized, livePreview, prompt, negativePrompt, negativeCapable, restoredSampler, samplerModelKey, samplerOptions, samplerChoice, width, height, steps, guidance, seed, batchSize, count, workflow, initImage, maskImage, strength, extendPct, extendSides, upscaleFactor, upscaleStrength, referenceImages, loras, loraCapable, controlnetCapable, controlnetId, controlImage, controlType, controlStrength, ensureSrc, loadGallery, refreshStatus, unifiedEdit, localizedMode, localizedLayer, maxExtras, referenceResolution, conditioning, editSize, editSizing, sizeLimits]);
+ }, [external, allowOversized, livePreview, prompt, negativePrompt, negativeCapable, restoredSampler, samplerModelKey, samplerOptions, samplerChoice, width, height, steps, guidance, seed, batchSize, count, workflow, initImage, maskImage, strength, extendPct, extendSides, upscaleFactor, upscaleStrength, referenceImages, loras, loraCapable, controlnetCapable, controlnetId, controlImage, controlType, controlStrength, ensureSrc, loadGallery, refreshStatus, unifiedEdit, localizedMode, localizedLayer, maxExtras, referenceResolution, conditioning, editSize, editSizing, sizeLimits]);
// Stop the in-flight generation. Latch FIRST, so a multi-run request stops even if the POST
// races the run that is already finishing.
@@ -4596,7 +4662,9 @@ export function ImagesPage({
const abort = new AbortController();
cancelAbort.current = abort;
try {
- const { cancelled } = await cancelDiffusionGeneration(abort.signal);
+ const { cancelled } = runTarget.current
+ ? await cancelConnectionImageGeneration(abort.signal)
+ : await cancelDiffusionGeneration(abort.signal);
cancelAcked.current = Boolean(cancelled);
if (!cancelled) setStopping(false);
} catch {
@@ -4616,6 +4684,10 @@ export function ImagesPage({
const handleGenerateWithRecall = useCallback(async () => {
if (busy !== null || !imagePresets.hydrated) return;
+ if (external) {
+ await handleGenerate();
+ return;
+ }
if (status?.loaded) {
const kind = status.model_kind;
if (
@@ -4672,6 +4744,7 @@ export function ImagesPage({
}, [
busy,
currentLoadAdvanced,
+ external,
handleGenerate,
handleLoad,
imagePresets.hydrated,
@@ -4721,6 +4794,10 @@ export function ImagesPage({
// Publish what the loaded model can do, so the sidebar submenu dims the rest. null while
// nothing is loaded, which leaves every workflow open to set up first.
useEffect(() => {
+ if (external) {
+ setSupported(["create"]);
+ return;
+ }
if (!status?.loaded) {
setSupported(null);
return;
@@ -4731,7 +4808,7 @@ export function ImagesPage({
wf.includes(t.requires === null ? "txt2img" : t.requires),
).map((t) => t.id),
);
- }, [status?.loaded, status?.workflows, setSupported]);
+ }, [external, status?.loaded, status?.workflows, setSupported]);
// Keep the active workflow valid for the loaded model: snap to the first supported one when capabilities change.
useEffect(() => {
@@ -4977,7 +5054,8 @@ export function ImagesPage({
<ModelSelector
triggerDataTour="images-model"
models={imageModels}
- value={selectorModelId}
+ externalModels={externalModels}
+ value={(external && externalModelId) || selectorModelId}
loadedModelIdOverride={selectorModelId}
activeGgufVariant={quant}
onValueChange={handleModelSelect}
@@ -5691,26 +5769,30 @@ export function ImagesPage({
{!unifiedEditActive && sizeControls}
{/* First of the one-line sliders, so it takes a bigger break than the gap gives. */}
- <div className="pt-2">
- <SliderField
- label="Steps"
- hint="Number of denoising steps. Start with the selected model's default; more steps take longer and may not improve quality."
- value={steps}
- min={1}
- max={50}
- step={1}
- onChange={setSteps}
- />
- </div>
- <SliderField
- label="Guidance"
- hint="Controls how strongly the model follows the prompt. Start with the selected model's default; distilled models may require low or zero guidance."
- value={guidance}
- min={0}
- max={15}
- step={0.5}
- onChange={setGuidance}
- />
+ {!external && (
+ <>
+ <div className="pt-2">
+ <SliderField
+ label="Steps"
+ hint="Number of denoising steps. Start with the selected model's default; more steps take longer and may not improve quality."
+ value={steps}
+ min={1}
+ max={50}
+ step={1}
+ onChange={setSteps}
+ />
+ </div>
+ <SliderField
+ label="Guidance"
+ hint="Controls how strongly the model follows the prompt. Start with the selected model's default; distilled models may require low or zero guidance."
+ value={guidance}
+ min={0}
+ max={15}
+ step={0.5}
+ onChange={setGuidance}
+ />
+ </>
+ )}
<SliderField
label="Batch size"
hint="How many images to make at once. Faster than running them one by one, but uses more VRAM. They share a seed but each one is different."
@@ -5773,21 +5855,25 @@ export function ImagesPage({
</div>
</Field>
)}
- <Field
- label="Seed"
- hint="Leave empty for a fresh random seed each run."
- className="pt-2"
- >
- <Input
- placeholder="Random if empty"
- value={seed}
- onChange={(e) => setSeed(e.target.value)}
- />
- </Field>
+ {!external && (
+ <>
+ <Field
+ label="Seed"
+ hint="Leave empty for a fresh random seed each run."
+ className="pt-2"
+ >
+ <Input
+ placeholder="Random if empty"
+ value={seed}
+ onChange={(e) => setSeed(e.target.value)}
+ />
+ </Field>
- <AdvancedDisclosure open={advancedOpen} onOpenChange={setAdvancedOpen}>
- {advancedControls}
- </AdvancedDisclosure>
+ <AdvancedDisclosure open={advancedOpen} onOpenChange={setAdvancedOpen}>
+ {advancedControls}
+ </AdvancedDisclosure>
+ </>
+ )}
</div>
{/* The scroll mask provides the fade; leave the footer unpainted to avoid dark-mode banding. */}
@@ -5808,11 +5894,11 @@ export function ImagesPage({
<Button
className="relative z-10 h-11 px-8 disabled:bg-muted disabled:text-muted-foreground disabled:opacity-100"
onClick={handleGenerateWithRecall}
- disabled={busy !== null || !imagePresets.hydrated || (!status?.loaded && !rememberedModel)}
+ disabled={busy !== null || !imagePresets.hydrated || (!status?.loaded && !rememberedModel && !external)}
>
Generate
</Button>
- {status?.loaded && (canReapply || status?.model_kind === "pipeline") && (
+ {!external && status?.loaded && (canReapply || status?.model_kind === "pipeline") && (
<Tooltip>
<TooltipTrigger asChild={true}>
<Button
@@ -6007,7 +6093,7 @@ export function ImagesPage({
{/* Same icon as the Images nav item. */}
<HugeiconsIcon icon={Image03Icon} className="size-12" strokeWidth={1.5} />
<p className="text-sm">
- {status?.loaded
+ {status?.loaded || external
? "Enter a prompt and hit Generate."
: "Select a diffusion model to load"}
</p>
diff --git a/studio/frontend/tests/image-generate-upstream-error.test.ts b/studio/frontend/tests/image-generate-upstream-error.test.ts
new file mode 100644
index 00000000000..52a569fb020
--- /dev/null
+++ b/studio/frontend/tests/image-generate-upstream-error.test.ts
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import { register } from "node:module";
+import test, { afterEach } from "node:test";
+
+register("./helpers/settings-api-resolver.mjs", import.meta.url);
+
+const originalFetch = globalThis.fetch;
+afterEach(() => {
+ globalThis.fetch = originalFetch;
+});
+
+const { cancelDiffusionGeneration, generateDiffusionImage, GenerateResponseLostError } = await import(
+ "../src/features/images/api.ts"
+);
+
+function respond(status: number, body: string, contentType: string) {
+ globalThis.fetch = (async () =>
+ new Response(body, { status, headers: { "content-type": contentType } })) as typeof fetch;
+}
+
+test("a JSON 502 from the app shows the upstream reason instead of waiting for a lost run", async () => {
+ respond(
+ 502,
+ JSON.stringify({ detail: "Image endpoint returned HTTP 401: invalid api key" }),
+ "application/json",
+ );
+ await assert.rejects(generateDiffusionImage({ prompt: "a sloth" }), (err: unknown) => {
+ assert.ok(!(err instanceof GenerateResponseLostError));
+ assert.match((err as Error).message, /invalid api key/);
+ return true;
+ });
+});
+
+test("a proxy's HTML 502 is still a lost response", async () => {
+ respond(502, "<html>Bad gateway</html>", "text/html");
+ await assert.rejects(
+ generateDiffusionImage({ prompt: "a sloth" }),
+ (err: unknown) => err instanceof GenerateResponseLostError,
+ );
+});
+
+test("Stop on a connection run targets the connection, a local Stop does not", async () => {
+ const urls: string[] = [];
+ globalThis.fetch = (async (input: RequestInfo | URL) => {
+ urls.push(String(input));
+ return new Response(JSON.stringify({ cancelled: true }), {
+ status: 200,
+ headers: { "content-type": "application/json" },
+ });
+ }) as typeof fetch;
+ const { cancelConnectionImageGeneration } = await import(
+ "../src/features/images/connection-cancel.ts"
+ );
+ await cancelConnectionImageGeneration();
+ await cancelDiffusionGeneration();
+ assert.deepEqual(urls, [
+ "/api/inference/images/generate/cancel?target=connection",
+ "/api/inference/images/generate/cancel",
+ ]);
+});
+
+test("a connected pick retires a staged local pick and skips the hidden seed", async () => {
+ const { readSrc } = await import("./helpers/kit.ts");
+ const page = readSrc("features/images/images-page.tsx");
+ const branch = page.slice(page.indexOf('if (meta.source === "external") {'));
+ const body = branch.slice(0, branch.indexOf("setExternalModelId(id);"));
+ assert.doesNotMatch(body, /downloadOnlyPick/);
+ assert.match(body, /pickGuard\.cancel\(\);/);
+ assert.match(body, /beginPick\(\);/);
+ assert.match(page, /if \(seed\.trim\(\) && !external\) \{/);
+});
+
+test("any local load takes the page back from a connection", async () => {
+ const { readSrc } = await import("./helpers/kit.ts");
+ const page = readSrc("features/images/images-page.tsx");
+ const load = page.slice(page.indexOf("const handleLoad = useCallback("));
+ const head = load.slice(0, load.indexOf("if (pollTimer.current) clearTimeout(pollTimer.current);"));
+ assert.match(head, /setExternalModelId\(null\);/);
+});
+
+test("the connection Stop stays out of the startup bundle's api module", async () => {
+ const { readSrc } = await import("./helpers/kit.ts");
+ assert.doesNotMatch(readSrc("features/images/api.ts"), /target=connection/);
+ assert.match(readSrc("features/images/images-page.tsx"), /from "\.\/connection-cancel"/);
+});
+
+test("a connected pick hides ControlNet without clearing the local setup", async () => {
+ const { readSrc } = await import("./helpers/kit.ts");
+ const page = readSrc("features/images/images-page.tsx");
+ const effect = page.slice(page.indexOf("setAvailableControlNets([]);") - 120);
+ assert.match(effect.slice(0, 160), /if \(!controlnetSupported\) \{/);
+ assert.match(page, /\}, \[controlnetSupported, status\?\.family\]\);/);
+});
diff --git a/studio/frontend/tests/image-model-recall.test.ts b/studio/frontend/tests/image-model-recall.test.ts
index 44e6ecf9bb5..39fd6ef8438 100644
--- a/studio/frontend/tests/image-model-recall.test.ts
+++ b/studio/frontend/tests/image-model-recall.test.ts
@@ -65,6 +65,7 @@ test("recalling a quantized model carries the selected adapters into its load",
vaeFile: "",
splitComponentFileList,
busy: null,
+ external: null,
imagePresets: { hydrated: true },
prompt: "a teapot",
pendingRecalledGeneration: { current: null },