Local ② engine: Qwen-Image 2.1 replaces Qwen-Image-Edit 2511 + angles LoRA (0.17.0)

.env.example

@@ -37,10 +37,9 @@ HF_TOKEN=
 # 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)
-# LDS_QWEN_EDIT_MODEL=qwen_image_edit_2511_int8_convrot.safetensors
-# LDS_ANGLES_LORA=qwen/Qwen-Image-Edit-2511-Multiple-Angles-LoRA.safetensors
-# LDS_QWEN_TEXT_ENCODER=qwen_2.5_vl_7b_fp8_scaled.safetensors
-# LDS_QWEN_VAE=qwen_image_vae.safetensors
+# LDS_QWEN21_MODEL=qwen_image_2.1_int8_convrot.safetensors
+# LDS_QWEN21_TEXT_ENCODER=qwen3vl_8b_int8_convrot.safetensors
+# LDS_QWEN21_VAE=qwen_image_2.1_vae_bf16.safetensors
 # LDS_UPSCALE_MODEL=4xNomosWebPhoto_RealPLKSR.safetensors
 # LDS_DEJPG_MODEL=1xDeJPG_OmniSR.pth
 # LDS_SAM3_CHECKPOINT=sam3.1_multiplex_fp16.safetensors

AGENTS.md

@@ -6,10 +6,10 @@ Turn a character, style, or concept into a ready-to-train LoRA dataset. One refe
 
 ## Current state
 
-_Last verified: 2026-08-23_
+_Last verified: 2026-10-03_
 
-- **Status:** in active development, released at v0.16.0 (git tag `v0.16.0`). CI green. **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.
+- **Status:** in active development, released at v0.16.0 (git tag `v0.16.0`); 0.17.0 (local ② generation moved to Qwen-Image 2.1) is committed but not yet tagged or released. CI green at 0.16.0. **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.
 - **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.
 - **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.
@@ -73,7 +73,7 @@ python cli.py keys --setup
 - ⏹ Stop is cooperative and checked *between* items; every stage calls `JOB.start()` first, and the Stop button must stay `queue=False`.
 - Bundled ComfyUI workflows are core-nodes-only, so an unknown node means an out-of-date ComfyUI — `comfy_api` preflights node classes and translates rejection bodies.
 - `tests/conftest.py` autouse fixture repoints the output roots at tmp; without it the suite writes real folders into `runs/`.
-- ② prompts never hard-code a medium — it comes from `studio/shot_style.py` (default: keep the reference's). Never write "photorealistic"/"hyperrealistic"; a test bans them from every generated prompt. Angle shots stay pure `<sks>` grammar.
+- ② prompts never hard-code a medium — it comes from `studio/shot_style.py` (default: keep the reference's). Never write "photorealistic"/"hyperrealistic"; a test bans them from every generated prompt. The local prompt never carries a negation ("without", "do not", a named prop): Qwen-Image 2.1 draws what it is told not to; a test bans it.
 - Re-value a Gradio dropdown with `gr.update(value=…)`, not `gr.Dropdown(value=…)` — the constructor form re-validates against the ORIGINAL `choices` and can silently drop the value.
 - The dataset **name and trigger are single header components** (`project_name`/`project_trigger`); every stage is wired straight to them. Never reintroduce a per-tab copy, and never mirror components per keystroke — Gradio's unqueued `.input` replies land out of order and settle on a prefix of what was typed.
 - Every model filename in a bundled ComfyUI template must be reachable from `comfy_api._MODEL_INPUTS` (plus a setting in `config.py` and a line in `.env.example`) — a test enforces it. Un-mapped means the user can't fix it in `.env` and `doctor` can't check it.

README.md

@@ -45,7 +45,7 @@ Run them in order (each step auto-fills the next) or jump straight to the one yo
 |---|---|---|
 | ① Restore / upscale | ComfyUI models, or basic Lanczos | — |
 | ① Subject isolation | **Built-in SAM3** (no ComfyUI) or ComfyUI SAM3 | — |
-| ② Generate shots *(character + concept)* | ComfyUI: Qwen Image Edit 2511 + Multiple-Angles LoRA | Gemini (Nano Banana) |
+| ② Generate shots *(character + concept)* | ComfyUI: Qwen-Image 2.1 *(Qwen Research License)* | Gemini (Nano Banana) |
 | ③ Caption | Qwen3-VL-8B, JoyCaption, NSFW finetune, **WD + e621 taggers**, LM Studio / Ollama / any OpenAI endpoint | Gemini Flash, Groq free tier |
 | ④ Export | always local (+ optional **.zip** and **Hugging Face** publish) | — |
 | ⑤ Train config | ai-toolkit (incl. SDXL) / **kohya sd-scripts** / musubi-tuner | — |
@@ -213,8 +213,7 @@ comic, digital illustration, traditional painting, 3D render, ink line art, or *
 where you describe the medium yourself (*"a 1970s screen-printed poster, halftone dots"*).
 
 Changing it rebuilds the shot plan, so hand-edited prompt cells are replaced — edit prompts
-after you've settled on a style. Camera-angle shots deliberately keep their bare
-turnaround grammar, which is what the angles LoRA was trained on. Use **👁 Preview final
+after you've settled on a style. Use **👁 Preview final
 prompt** to see exactly what a row will send, including outfit and prop-exclusion. The style
 is recorded in the exported `metadata.json` and picked up by ⑤'s sample prompt.
 

app.py

@@ -54,7 +54,7 @@
 
 ENGINE_CHOICES = [
     ("Cloud — Gemini image model (best identity fidelity, SFW only)", "gemini"),
-    ("Local — ComfyUI Qwen Image Edit 2511 (free, private, uncensored)", "comfyui"),
+    ("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()]
 CAPTIONER_CHOICES = [(c.label, c.key) for c in CAPTIONERS]
@@ -303,7 +303,7 @@ def preview_final_prompt(plan_df: pd.DataFrame, engine: str, exclude_props: bool
     if exclude_props:
         shot = apply_prop_exclusion(shot)
     field = "local_prompt" if engine == "comfyui" else "cloud_prompt"
-    which = "Local (ComfyUI / Qwen-Edit)" if engine == "comfyui" else "Cloud (Gemini)"
+    which = "Local (ComfyUI / Qwen-Image 2.1)" if engine == "comfyui" else "Cloud (Gemini)"
     return (f"**{which} prompt for `{shot.id}`** — exactly what the engine receives, "
             f"after outfit and prop-exclusion are folded in:\n\n```\n"
             f"{getattr(shot, field)}\n```")
@@ -1343,7 +1343,15 @@ def do_load_plan(plan_name: str):
         raise gr.Error(f"Couldn't read the plan at {path}: {e}. Check the YAML — "
                        f"each shot needs at least an `id`, `kind`, `local_prompt` "
                        f"and `cloud_prompt`.") from e
-    return _shots_to_df(shots), f"✅ Loaded {len(shots)} shots from {path}"
+    note = f"✅ Loaded {len(shots)} shots from {path}"
+    # Plans saved before 0.17.0 carry the old engine's `<sks>` LoRA grammar, which
+    # Qwen-Image 2.1 reads as literal text and renders badly.
+    stale = sum("<sks>" in s.local_prompt for s in shots)
+    if stale:
+        note += (f" ⚠️ {stale} shot(s) still use the old Qwen-Image-Edit `<sks>` prompts, "
+                 f"which the local Qwen-Image 2.1 engine doesn't understand — rebuild the "
+                 f"plan from the dataset type / shot style, or rewrite those local prompts.")
+    return _shots_to_df(shots), note
 
 
 def estimate_cost(engine: str, cloud_model: str, df: pd.DataFrame) -> str:
@@ -1795,15 +1803,18 @@ def _check_for_update():
                     gen_exclude_props = gr.Checkbox(
                         value=True,
                         label="Exclude props/accessories from the reference",
-                        info="Asks the generator to drop bags, held objects and "
-                             "accessories carried in your reference, so they don't get "
-                             "baked into every dataset image. Isolating the source in ① "
-                             "is the more reliable fix. Character-oriented wording — off "
+                        info="Cloud engine only: asks Gemini to drop bags, held objects "
+                             "and accessories carried in your reference, so they don't "
+                             "get baked into every dataset image. The local engine "
+                             "ignores it (naming a prop, even to forbid it, makes "
+                             "Qwen draw it) — isolate the source in ① instead, the more "
+                             "reliable fix either way. Character-oriented wording — off "
                              "by default for Concept datasets.")
                     gen_isolate = gr.Checkbox(value=False,
                                               label="Isolate generated angle shots (white background)",
                                               info="Cut generated angle shots onto white too "
-                                                   "(helps the angles LoRA on back views).")
+                                                   "(replaces each angle shot's setting "
+                                                   "with a plain white background).")
                     gen_iso_backend = gr.Dropdown(ISOLATION_CHOICES,
                                                   value=settings.isolation_backend,
                                                   label="Isolation backend",