User scope: all25 supplied images as-is in WORKSTATION/new_lora, with the
literal tag susanah; remove their visible watermark overlays first, preserve dimensions
and image quality, then train and deliver15 generated samples through Hermes Telegram.
The source contains multiple women; the user explicitly confirmed using the whole set,
so this adapter learns a mixed appearance rather than one verified person’s identity.
All25 images are unique decoded images. Source JPEGs are untouched. Cleaned lossless PNG
copies are in WORKSTATION/cleaned; per-image records and masks are under
cleaning/. DeepSeek vision located corner overlays, then the masks were visually checked.
The SAME Qwen Image2.1 checkpoint used for generation performs reference-based editing.
The first three images used whole-frame inference; later marks used local context crops
to preserve native detail more efficiently. Only the masked pixels were composited back.
Every output retains exact original dimensions and exact decoded pixels outside its mask.
The sky repair in image16 was repeated with a tighter mask and12px inward feather.
Captions describe outfit, pose, expression and setting and include susanah literally.
Fresh Qwen2.1 rank/alpha32, LR1e-4, AdamW8bit,600 completed steps,768 buckets, cached text
and latents, GPU0. Checkpoints and fixed-seed samples every200 support model selection.
No Lucianax/Photo15x adapter is used as initialization. All GPUs retain245W power limits.
The final step600 checkpoint was selected after visual comparison against200 and400.
Both control scenes retain requested outfits; the final second sample follows the requested
smile more clearly and has a more natural pose than400. Full-body framing still crops feet.
These two subjective comparisons do not establish general superiority or identity consistency.
Audit:384 finite tensors,192 nonzero LoRA B matrices; trainer exited0. Training loop24m13s,
approximately2.42s/step, excluding caching, loading and preview overhead.
Selected SHA-256: f5e2a6ac1bbb6b1045a52e7df119827fdb01af3497a8eb4c3f08c52b7ab68e05.
Canonical alias: WORKSTATION/susanah_qwen_image_21_lora_v1.safetensors.
The200/400 checkpoints remain available for comparison.
The15 test prompts are expanded by the requested PE-T2I MLX8bit model on the MacBook, then rendered locally with the selected Susanah LoRA at1024x1536,25 steps, CFG1, strength1.0, seeds3163501–3163515. Actual delivery receipts will be added after review.
Frontend: private workstation service (Tailscale), choose profile susanah.
Editable ComfyUI JSON: WORKSTATION/susanah.json.
API JSON: WORKSTATION/susanah.json.
The shared profile supports CLI/MCP generation with or without LoRA; no disabled MCP
registration was enabled.13 integration tests pass and both LoRA/base graph routes were checked.
Status: DONE. All25 cleaned copies audited, training completed and15 samples delivered. Images, exact captions/prompts and weights remain outside this repository.
Completed delivery:15 visually reviewed PNG samples,1024x1536, plus original PNG ZIP. Generation mean26.02s, range26.01–26.02s, total390.29s (prompt enhancement measured separately). Telegram receipts: 12761, 12762, 12763, 12764, 12765, 12766, 12767, 12768, 12769, 12770, 12771, 12772, 12773, 12774, 12775, 12776. Delivery used the user-authorized Hermes channel skill. All source originals remain intact. Local sample metadata retains seed, exact workflow, selected adapter SHA and output SHA. All three GPUs remain at245W.