Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,64 @@
+---
+license: other
+base_model: black-forest-labs/FLUX.2-klein-base-9B
+tags:
+  - flux
+  - flux-2
+  - flux-klein
+  - lora
+  - product-render
+  - air-conditioner
+  - hvac
+---
+
+# Toshiba AC Wall Unit — Flux 2 Klein 9B LoRA (v7)
+
+A LoRA fine-tuned on **FLUX.2-klein-base-9B** to generate consistent commercial product renders of a Toshiba wall-mounted air conditioning unit.
+
+## Usage
+
+**Trigger word:** `uctosh_1hvac`
+
+**Base model:** `black-forest-labs/FLUX.2-klein-base-9B` (must use Klein 9B, not Dev or Schnell)
+
+**Recommended LoRA strength:** 0.5–0.75 (start at 0.65)
+
+### Prompt examples
+
+Product render:
+```
+uctosh_1hvac, a wall-mounted climate unit, louvres closed, front view, product render on white background
+```
+
+Lifestyle scene:
+```
+uctosh_1hvac, a wall-mounted climate unit, louvres open, mounted high on a white wall in a modern living room, soft natural daylight
+```
+
+Add `no text, no logos, no branding` to suppress text artifacts.
+
+## Training Details
+
+| Parameter | Value |
+|---|---|
+| Base Model | FLUX.2-klein-base-9B |
+| Text Encoder | Qwen3 8B (quantized) |
+| Rank | 128 |
+| Alpha | 128 |
+| Steps | 2000 (checkpoint at 1500 selected) |
+| Learning Rate | 9.5e-5 |
+| Optimizer | AdamW8bit |
+| Dataset | 56 images (50 angles + 6 close-ups) |
+| Caption Strategy | Exclusion principle — minimal captions |
+| Toolkit | Ostris AI Toolkit |
+| GPU | NVIDIA L40S 48GB |
+
+## Checkpoint
+
+This is the **step 1500** checkpoint — selected as the optimal balance of quality vs overfitting. Steps beyond 1750 showed overfitting (texture bleeding into scene objects).
+
+## fal.ai Inference
+
+Use endpoint: `fal-ai/flux-2/klein/9b/base/lora`
+
+LoRA URL: `https://huggingface.co/IanHancock/uctosh-1hvac-flux2-klein-9b-lora-v7/resolve/main/toshiba_ac_v7_000001500.safetensors`