EnragedAntelope · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026
diff --git a/.env.example b/.env.example
@@ -36,7 +36,7 @@ HF_TOKEN=
 # reserved by Windows (Hyper-V/WSL/Docker). Set it to pin one exactly.
 # LDS_PORT=7861
 # LDS_OUTPUT_ROOT=./datasets
-# LDS_GEMINI_IMAGE_MODEL=gemini-3-pro-image-preview
+# LDS_GEMINI_IMAGE_MODEL=auto          # auto = best image model your key can use; or pin an id from ② refresh
 # LDS_UPDATE_CHECK_ENABLED=true         # checks GitHub releases on launch (no data sent); set false to disable
 
 # ComfyUI model filenames, if yours are named differently (see docs/comfyui-setup.md)

diff --git a/AGENTS.md b/AGENTS.md
@@ -6,12 +6,12 @@ Turn a character, style, or concept into a ready-to-train LoRA dataset. One refe
 
 ## Current state
 
-_Last verified: 2026-10-03_
+_Last verified: 2026-10-08_
 
-- **Status:** in active development, released at v0.17.2 (git tag `v0.17.2`). **Every version bump must ship a GitHub Release** or the in-app update check never fires. The project was renamed twice (lora-dataset-studio → lora-distillery → dataset-deviser); older references under the previous names are stale.
-- **Works:** all five stages end to end (preprocess → generate & curate → caption → export → train config); the three dataset types (character, style, concept) with their own shot plans and caption framing; local and cloud paths for every stage; gallery curation with selection carried forward between stages, including shift-click range select; advisory dedupe, quality flags and caption lint; trainer configs for ai-toolkit (incl. SDXL), kohya sd-scripts and musubi-tuner; opt-in private Hugging Face publish. Since 0.15.0: per-image fault isolation in ① (one bad source never ends the batch), a cooperative ⏹ Stop across ①/②/③ with documented resume paths, translated ComfyUI failures (missing node, bad input, unreachable server) and an in-app 🩺 setup check. Since 0.15.1: a ② **shot style** (default = keep the reference's own medium; eight presets + custom text) threaded into both prompt builders, the CLI, ④ metadata and ⑤'s sample prompt, a final-prompt preview, and a prev/next/save-and-next caption editor with the image on screen. Since 0.16.0: **one name box and one trigger box in the header** that ②/③/④/⑤ all read (they used to be per-tab and went stale), ④ stating the pair it stamped, an opt-in ④ **hand-off to [Idiot LoRa Builder](https://github.com/Fablestarexpanse/Idiot-Lora-Builder)** (writes its `.lora-studio/ratings.json` so its grid opens pre-triaged — nothing is launched), a ⑤ advisory when the dataset can't fill the training resolution, every bundled ComfyUI model filename overridable from `.env`, and a `cli build` preflight for the local engine. Since 0.17.0: the local ② engine is **Qwen-Image 2.1** (plain-English prompts shared with Gemini, ~26 s/shot, `LDS_QWEN21_*` filenames) instead of Qwen-Image-Edit 2511 + the Multiple-Angles LoRA, whose graph stays in `studio/comfy_workflows/legacy/`; a dead ComfyUI mid-batch now fails one shot, not the run. Since 0.17.1: the UI port is chosen by `_pick_port` in `app.py` — 7861 if bindable, else an OS-assigned free port (Windows reserves a ~500-port block around 7861), or `LDS_PORT` to pin one. Since 0.17.2: `comfy_api.upload_image` downscales a source over 90 MB before upload (ComfyUI answers 413 above its 100 MB `--max-upload-size`).
-- **In progress:** nothing half-built — 0.16.0 closed out the hand-off, cross-tab-identity and config-hardening work. `docs/ARCHITECTURE.md` carries the "Future ideas" and "Deferred" sections that hold the real backlog.
-- **Known gaps / next steps:** pick up from `docs/ARCHITECTURE.md` → "Future ideas" / "Deferred", and read its *Maintainer principles* before changing anything; `gradio` is pinned `<6` because of a real stuck-loading-overlay regression — unpinning needs that verified fixed upstream; training is never launched for you and is not planned to be; the default export `output_root` writes into the repo root, so a smoke-test export leaves an untracked dataset folder to delete before committing.
+- **Status:** in active development; main is v0.17.3 (git tag `v0.17.3`). **Every version bump must ship a GitHub Release** or the in-app update check never fires. The project was renamed twice (lora-dataset-studio → lora-distillery → dataset-deviser); older references under the previous names are stale.
+- **Works:** all five stages end to end (preprocess → generate & curate → caption → export → train config); the three dataset types (character, style, concept) with their own shot plans and caption framing; local and cloud paths for every stage; gallery curation with selection carried forward between stages, including shift-click range select; advisory dedupe, quality flags and caption lint; trainer configs for ai-toolkit (incl. SDXL), kohya sd-scripts and musubi-tuner; opt-in private Hugging Face publish. Since 0.15.0: per-image fault isolation in ① (one bad source never ends the batch), a cooperative ⏹ Stop across ①/②/③ with documented resume paths, translated ComfyUI failures (missing node, bad input, unreachable server) and an in-app 🩺 setup check. Since 0.15.1: a ② **shot style** (default = keep the reference's own medium; eight presets + custom text) threaded into both prompt builders, the CLI, ④ metadata and ⑤'s sample prompt, a final-prompt preview, and a prev/next/save-and-next caption editor with the image on screen. Since 0.16.0: **one name box and one trigger box in the header** that ②/③/④/⑤ all read (they used to be per-tab and went stale), ④ stating the pair it stamped, an opt-in ④ **hand-off to [Idiot LoRa Builder](https://github.com/Fablestarexpanse/Idiot-Lora-Builder)** (writes its `.lora-studio/ratings.json` so its grid opens pre-triaged — nothing is launched), a ⑤ advisory when the dataset can't fill the training resolution, every bundled ComfyUI model filename overridable from `.env`, and a `cli build` preflight for the local engine. Since 0.17.0: the local ② engine is **Qwen-Image 2.1** (plain-English prompts shared with Gemini, ~26 s/shot, `LDS_QWEN21_*` filenames) instead of Qwen-Image-Edit 2511 + the Multiple-Angles LoRA, whose graph stays in `studio/comfy_workflows/legacy/`; a dead ComfyUI mid-batch now fails one shot, not the run. Since 0.17.1: the UI port is chosen by `_pick_port` in `app.py` — 7861 if bindable, else an OS-assigned free port (Windows reserves a ~500-port block around 7861), or `LDS_PORT` to pin one. Since 0.17.2: `comfy_api.upload_image` downscales a source over 90 MB before upload (ComfyUI answers 413 above its 100 MB `--max-upload-size`). Since 0.17.3: the cloud image model defaults to **`auto`** (Google shut down the `-preview` ids on 2026-06-25; `resolve_image_model` picks from the cached live list, Nano Banana 2.1 first), Groq captions use `qwen/qwen3.8-27b` (key `groq-qwen`, old key aliased), every API image goes through `config.api_image` (true mime, size cap), musubi commands include the cache steps and the per-architecture network module, ai-toolkit gains Qwen-Image 2.1 and a correct Krea 2 (Raw) preset and keeps every checkpoint, ① applies EXIF rotation, no longer 4x-upscales large JPEGs, and can jump a busy ComfyUI queue like ② (`--front` on `preprocess`/`build`).
+- **In progress:** a phased LoRA best-practice update. 0.18.0 is next: epochs as the ⑤ unit with an exposure line, a validation pack, held-out references, a Fizgig trainer target, source-size provenance in a PNG chunk. Then 0.19.0 (training copies keep their background, identity policy, framing tiers, coverage panel, local multi-reference) and 0.20.0 (a Quick build tab). Nothing from those is half-built on main.
+- **Known gaps / next steps:** Auto's model order (`IMAGE_MODEL_PREFERENCE`, Nano Banana 2.1 then Pro) rests on one A/B subject — see ARCHITECTURE.md; local ② still sends only the first reference. Otherwise pick up from `docs/ARCHITECTURE.md` → "Future ideas" / "Deferred", and read its *Maintainer principles* before changing anything; `gradio` is pinned `<6` because of a real stuck-loading-overlay regression — unpinning needs that verified fixed upstream; training is never launched for you and is not planned to be; the default export `output_root` writes into the repo root, so a smoke-test export leaves an untracked dataset folder to delete before committing.
 - **Deep docs:** `docs/ARCHITECTURE.md` (module map, stage and backend detail, gotchas, backlog — the deep reference), `docs/comfyui-setup.md`. Worklogs under `docs/` are gitignored and local-only by design.
 
 ## Architecture in 60 seconds

diff --git a/README.md b/README.md
@@ -201,8 +201,8 @@ drop-list. It's advisory: nothing is blocked or rewritten.
 | Qwen3-VL-8B NSFW-Caption | your GPU, ~17 GB | explicit-dataset specialist |
 | **WD EVA02 / ViT v3** *(tagger)* | GPU or CPU, ONNX | canonical **Danbooru** tags |
 | **Z3D e621 ConvNeXt** *(tagger)* | GPU or CPU, ONNX | canonical **e621/furry** tags |
-| Gemini Flash | Google API | SFW, ~$0.0007/img est., billed to your key |
-| Groq Qwen3.6 27B | Groq API | SFW, free tier, rate-limited |
+| Gemini Flash | Google API | SFW, ~$0.002/img est., billed to your key |
+| Groq Qwen3.8 27B | Groq API | SFW, free or paid plan, rate-limited |
 | LM Studio / Ollama / custom | your choice | any OpenAI-compatible endpoint |
 
 ## Shot style

diff --git a/app.py b/app.py
@@ -40,13 +40,19 @@
 from studio.config import (
     CAPTIONERS,
     CAPTIONERS_BY_KEY,
-    CLOUD_IMAGE_PRICES,
     friendly_api_error,
     list_images,
     load_caption_model_cache,
     read_caption,
     settings,
 )
+from studio.engines.gemini import (
+    AUTO_MODEL,
+    image_model_choices,
+    image_price,
+    known_image_models,
+    resolve_image_model,
+)
 from studio.jobs import JobControl
 from studio.shotplan import Shot, apply_prop_exclusion, apply_wardrobe, plan_for_type
 from studio.trainer_configs import TRAINER_MODELS, TRAINERS
@@ -57,7 +63,17 @@
     ("Cloud — Gemini image model (best identity fidelity, SFW only)", "gemini"),
     ("Local — ComfyUI Qwen Image 2.1 (free, private, uncensored)", "comfyui"),
 ]
-CLOUD_MODEL_CHOICES = [(f"{m}  (~${p:.3f}/img est.)", m) for m, p in CLOUD_IMAGE_PRICES.items()]
+CLOUD_MODEL_CHOICES = image_model_choices(known_image_models())
+
+
+def _cloud_model_default(pinned: str, choices: list[tuple[str, str]]) -> str:
+    """The ② dropdown's start value: the .env pin (retired ids mapped) if listed, else Auto."""
+    if pinned != AUTO_MODEL:
+        pinned = resolve_image_model(pinned)
+    return pinned if pinned in {m for _, m in choices} else AUTO_MODEL
+
+
+CLOUD_MODEL_DEFAULT = _cloud_model_default(settings.gemini_image_model, CLOUD_MODEL_CHOICES)
 CAPTIONER_CHOICES = [(c.label, c.key) for c in CAPTIONERS]
 
 # Gemini caption-model dropdown seed: use the local cache if present, else a
@@ -581,7 +597,8 @@ def _preprocess_note(reports, out_dir: Path, alpha_cutout: bool) -> str:
 def do_preprocess(files: list[str], folder: str, target: int, restore_mode: str,
                   restore_backend: str, isolate: bool, isolation_backend: str,
                   subject_prompt: str, exclude_prompt: str, tighten: bool = False,
-                  alpha_cutout: bool = False, progress=gr.Progress()):
+                  alpha_cutout: bool = False, front: bool = False,
+                  progress=gr.Progress()):
     sources = _inputs(files, folder)
     out_dir = _stamped("prepped")
     force = {"Auto (only if needed)": None, "Always": True, "Never": False}[restore_mode]
@@ -598,7 +615,7 @@ def report(msg: str):
             subject_prompt=subject_prompt or "character",
             exclude_prompt=exclude_prompt or "", restore_backend=restore_backend,
             isolation_backend=isolation_backend, tighten_crop=tighten,
-            alpha_cutout=alpha_cutout, should_stop=JOB, progress=report)
+            alpha_cutout=alpha_cutout, front=front, should_stop=JOB, progress=report)
     except OSError as e:
         raise gr.Error(f"Couldn't write to '{out_dir}': {e}. Check the output folder "
                        f"(valid drive, writable, enough space).") from e
@@ -1358,17 +1375,10 @@ def do_load_plan(plan_name: str):
 def estimate_cost(engine: str, cloud_model: str, df: pd.DataFrame) -> str:
     n = len(df)
     if engine == "gemini":
-        from studio.config import CLOUD_IMAGE_PRICES, load_cloud_model_cache
-
-        price = CLOUD_IMAGE_PRICES.get(cloud_model)
-        cached = load_cloud_model_cache() or []
-        for m in cached:
-            if m.get("model_id") == cloud_model and m.get("price") is not None:
-                price = m["price"]
-                break
+        model, price = image_price(cloud_model)
         if price is None:
-            return f"**Cost:** {n} images on `{cloud_model}` (price unknown — billed to your API key)"
-        return (f"**Cost:** ~${n * price:.2f} for {n} images on `{cloud_model}` "
+            return f"**Cost:** {n} images on `{model}` (price unknown — billed to your API key)"
+        return (f"**Cost:** ~${n * price:.2f} for {n} images on `{model}` "
                 f"(estimate at build time — billed to your own Google API key)")
     return f"**Cost:** {n} images, $0 (local generation)"
 
@@ -1534,15 +1544,16 @@ def do_generate_train_config(trainer: str, model_key: str, dataset_dir: str,
             f"trainer's own docs before a long run.")
 
 
-def refresh_cloud_models(force: bool = False):
+def refresh_cloud_models(current: str = "", force: bool = False):
+    """New choices from the API; keeps the user's pick when it is still offered."""
     from studio.engines.gemini import list_image_models
 
     try:
-        models = list_image_models(force_refresh=force)
+        choices = image_model_choices(list_image_models(force_refresh=force))
     except Exception as e:
         raise gr.Error(f"Could not list models: {e}") from e
-    return gr.Dropdown(choices=models,
-                       value=models[0][1] if models else settings.gemini_image_model)
+    keep = current if current in {m for _, m in choices} else AUTO_MODEL
+    return gr.Dropdown(choices=choices, value=keep)
 
 
 def refresh_caption_models():
@@ -1749,6 +1760,10 @@ def _check_for_update():
                              "own compositing workflows. Builtin SAM3 backend only. Leave off "
                              "(default) if you're continuing to ② Generate — it expects a white "
                              "background reference. Needs isolation on.")
+                    pre_front = gr.Checkbox(
+                        value=False, label="Prioritize this app's ComfyUI jobs",
+                        info="Puts ComfyUI restore/isolation jobs at the head of its "
+                             "pending queue. Does not interrupt a job already running.")
                     btn_pre = gr.Button("① Preprocess", variant="primary")
                 with gr.Column(scale=2):
                     pre_note = gr.Markdown()
@@ -1793,8 +1808,10 @@ def _check_for_update():
                                       label="Generation engine",
                                       info="Cloud Gemini needs no GPU (best identity, SFW); "
                                            "local ComfyUI is free, private, uncensored.")
+                    # A pinned id the list lacks (a retired preview in .env) would
+                    # make Gradio reject every event that reads this dropdown.
                     cloud_model = gr.Dropdown(CLOUD_MODEL_CHOICES,
-                                              value=settings.gemini_image_model,
+                                              value=CLOUD_MODEL_DEFAULT,
                                               label="Cloud image model",
                                               info="Only used by the Cloud engine. Prices are "
                                                    "build-time estimates.")
@@ -2200,7 +2217,7 @@ def _check_for_update():
         do_preprocess,
         [pre_files, pre_folder, target, restore_mode, restore_backend, isolate,
          isolation_backend, subject_prompt, exclude_prompt, pre_tighten,
-         pre_alpha_cutout],
+         pre_alpha_cutout, pre_front],
         [prep_gallery, pre_note, log_box, gen_src_folder, cap_folder]) \
            .then(lambda s, e: (s, e), [subject_prompt, exclude_prompt],
                  [gen_subject, gen_exclude])
@@ -2249,12 +2266,12 @@ def _check_for_update():
     btn_outfits_clear.click(clear_outfits, [plan], [plan, plan_note])
     btn_save_plan.click(do_save_plan, [plan, plan_name], [plan_note])
     btn_load_plan.click(do_load_plan, [plan_name], [plan, plan_note])
-    refresh_models.click(refresh_cloud_models, [], [cloud_model])
+    refresh_models.click(refresh_cloud_models, [cloud_model], [cloud_model])
 
-    def _force_refresh():
-        return refresh_cloud_models(force=True)
+    def _force_refresh(current: str):
+        return refresh_cloud_models(current, force=True)
 
-    force_refresh_models.click(_force_refresh, [], [cloud_model])
+    force_refresh_models.click(_force_refresh, [cloud_model], [cloud_model])
     engine.change(estimate_cost, [engine, cloud_model, plan], [cost])
     cloud_model.change(estimate_cost, [engine, cloud_model, plan], [cost])
     plan.change(estimate_cost, [engine, cloud_model, plan], [cost])