0.18.0: epochs, validation pack, held-out photos, Fizgig target

AGENTS.md

@@ -8,9 +8,9 @@ Turn a character, style, or concept into a ready-to-train LoRA dataset. One refe
 
 _Last verified: 2026-10-08_
 
-- **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.
+- **Status:** in active development; main is v0.18.0 (git tag `v0.18.0`). **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, musubi-tuner and Fizgig; 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`). Since 0.18.0: ⑤ counts **epochs** (one checkpoint and one sample set each) with auto repeats sized to a step target and a live exposure line; every config ships a `validation/` pack (8 fixed prompts at seed 42, `validation.md`, a per-epoch score sheet); **Fizgig** is a ⑤ target (Krea 2 / Qwen Image 2.1 / Klein); ① stamps a `dd_source` provenance PNG chunk so ⑤ names upscaled sources and ④ knows real photos; ④ can hold out N real photos in a `-heldout` sibling as a likeness test.
+- **In progress:** a phased LoRA best-practice update. 0.19.0 is next: training copies keep their background, identity policy (outfits vary by default), framing tiers, a coverage panel, a full-body warning, near-duplicate/burst flags, an anchor shot, local multi-reference. Then 0.20.0 (a Quick build tab). Nothing from those is half-built on main. The 0.18.0 optimizer dropdown and attention-only toggle were deferred (ARCHITECTURE.md → Future ideas).
 - **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.
 
@@ -21,7 +21,7 @@ _Last verified: 2026-10-08_
 - **Local OR cloud per stage.** Every step has a free/private local path AND a no-GPU cloud path. Mix and match: generate on cloud, caption on GPU, or reverse.
 - **Multiple backends:** ComfyUI (generation, preprocessing), built-in SAM3 (isolation), ONNX booru tagger (WD + e621), Qwen3-VL-8B / JoyCaption / NSFW finetune (local captioning), Gemini Flash / Groq (cloud captioning), LM Studio / Ollama / any OpenAI-compatible endpoint.
 - **Curate by clicking.** Toggle shots from the gallery with ✅/⬜ marks. Selection carries through stages: what you keep in ② is what ③ captions, and what ③ captions is what ④ preselects.
-- **Trainer configs:** Generated for ai-toolkit (incl. SDXL), kohya sd-scripts, musubi-tuner. Training is NEVER launched for you — configs are shown, you run them.
+- **Trainer configs:** Generated for ai-toolkit (incl. SDXL), kohya sd-scripts, musubi-tuner, Fizgig, plus a validation pack. Training is NEVER launched for you — configs are shown, you run them.
 - **Honest by design.** No telemetry, no upsell. Cloud calls bill YOUR key; this tool takes no cut.
 
 ## Layout

README.md

@@ -6,7 +6,7 @@ settings; for a **concept**, 18 shots around one object — a turnaround plus fr
 For a **style**, bring your own images and get smart, correctly-framed captions. Either way the
 output is a flat folder that drops straight into any trainer
 (**ai-toolkit / kohya / OneTrainer / …**), plus a ready-to-edit **training config** for
-ai-toolkit, kohya or musubi.
+ai-toolkit, kohya, musubi or Fizgig, and a validation pack for picking the best checkpoint.
 
 ![Generate & curate tab](docs/images/ui-generate.png)
 
@@ -48,7 +48,7 @@ Run them in order (each step auto-fills the next) or jump straight to the one yo
 | 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 | — |
+| ⑤ Train config | ai-toolkit (incl. SDXL) / **kohya sd-scripts** / musubi-tuner / **Fizgig** | — |
 
 ## Dataset types
 
@@ -230,6 +230,14 @@ Both have a switch in ②, because both bit us:
   *is* the character). Best fix: **isolate the source in ①** — SAM3 keeps bags and held objects
   out of the subject mask. The "Exclude props" toggle in ② helps too, especially on Gemini.
 
+## Picking the best checkpoint
+
+⑤ sets training in **epochs**: every epoch saves a checkpoint and renders the same 8 test
+prompts at the same seed. Leave **Repeats** at 0 and it sizes each epoch for you; the exposure
+line shows the total. The `validation/` folder it writes holds the prompts, a short guide and a
+score sheet. To test likeness fairly, set **Hold out N reference photos** in ④: those photos
+stay out of training, in a `-heldout` folder beside the dataset, to compare samples against.
+
 ## CLI
 
 Every stage is a standalone subcommand; `--help` shows all options.