Update README.md

README.md

@@ -48,24 +48,28 @@ pipeline_tag: text-to-image
   - Training Tools: [Onetrainer (Koratahiu‘s Fork)](https://github.com/Koratahiu/OneTrainer/tree/moss-image) / [Fizgig](https://github.com/shootthesound/Fizgig) / [Fizgig-Advanced(Murf-Code's Fork)](https://github.com/Murf-Code/Fizgig-Advanced)
   - Dataset: Mostly 50+ images (averaging 100+)
   - Tagging Style: Mixed format (Booru tags as prefixes combined with natural language)
-  - Training Res: 512 / 0.25(0.26) MP 
+  - Training Res: 512 / 0.25(0.26) MP
   - Repeats per Image: 50+
   - Training Parameters: Default settings from Onetrainer / Fizgig
     - Onetrainer tweak:
       - Attention:`torch CuDNN` (overriding the default SDPA)
+      - LR:`1.5e-04 - 2.5e-04`
+      - Batch Size:：`3 - 4`
     - Fizgig tweak:
       - Presets: `Moss-Image Fast(rank 16, 0.25MP)`
+      - LR:`1e-04`
       - Pictrue Size: `0.26 MP` ≈ 512 res
       - Gradient Accumulation: `3`
      - Fizgig-Advanced tweak:
       - Presets: `Moss-Image Fast(rank 16, 0.25MP)`
       - Pictrue Size: `0.26 MP` ≈ 512 res
-      - Batch Size: `3`
+      - LR:`1.5e-04 - 2.5e-04`
+      - Batch Size: `3 - 4`
   - Training Speed (Benchmark: 39 imgs, 0.26MP, r16):
     - OneTrainer: 910 steps, 1.12 s/it (`Local Batch Size: 3, LR: 0.0003`) → ~0.37s/img (17m 43s)
     - Fizgig(Batch-Size: Not Supported): 2730 steps, 1.25 it/s (`Gradient Accumulation: 3, LR: 0.0001`) → ~0.83s/img (36m 17s)
     - Fizgig-Advanced(Batch-Size: Supported): 910 steps, 2.07 s/it (`Batch Size: 3, LR: 0.0003`) → ~0.69s/img (31m 20s)
-    - *(Note on Speed & Formats: OneTrainer used the official sharded Diffusers format (bfloat16), while Fizgig used the ComfyUI single-file format (int8). Although int8/fp8 usually compute faster on modern GPUs, the benchmark shows Fizgig is slower here—likely due to software-level framework overheads. Fizgig also supports the original bf16 Diffusers shards.)*
+    - *(Note on Speed & Formats: OneTrainer used the official sharded Diffusers format (bfloat16), while Fizgig used the ComfyUI single-file int8 format.)*
   - Showcase & Card Image Parameters:
     - Model: - 
     - LoRA Weight: 0.8 ~ 1