v1.3.1

[1.3.1]

Fixed

• LoRA Bake-in crashed on LoRAs containing convolutional layers with non-1×1 kernels (RuntimeError: Expected size for first two dimensions of batch2 tensor to be [...]). The matmul reconstruction now handles Conv2d layers with both the standard kohya convention (down: [rank, in_ch, kh, kw]) and the inverted-rank variant some trainers produce.
• Individual layers whose shape can't be reconciled are now skipped with a ⚠️ Skipped <layer> warning instead of aborting the whole bake — the rest of the LoRA still fuses.

v1.3.0

[1.3.0]

Added

• Merge & Gen button on the Merge tab — runs the merge, loads the merged model into the WebUI, and generates a preview image using the prompt and settings from the txt2img tab. The result is shown in an accordion inside NeoMerger without needing to switch tabs. Generation runs without txt2img extensions (ADetailer, HiRes Fix, etc.) — it's a clean preview of the merged base model; use the txt2img tab for refined generation.
• Convert to Normal button in Easy mode — expands the semantic sliders into the 20 per-block weights and switches to Normal mode for fine-tuning. No loss of fidelity — uses the same calculation the merge engine applies internally.
• "Include experimental sliders" checkbox in the Experimental accordion — when off (default), experimental sliders are skipped entirely from the weighted-average calculation so they don't dilute the main category values. State is saved with presets.
• Base Alpha slider in the no-blocks panel — controls the base block (text encoder · CLIP) independently from the global Alpha which is applied to the UNet. Especially useful for TIES and DARE methods, which don't support per-block weights. The value is also saved with presets and embedded in checkpoint metadata.
• JavaScript helper at javascript/neomerger.js to read txt2img fields for Merge & Gen.
  Changed
• Default behavior of Easy mode no longer dilutes block weights with experimental categories. Existing presets keep their old behavior on load (toggle defaults to off, matching how those presets were used before).
• Fixed Lora Bake-in for Anima models, now fuses all the layers
  Notes
• Old presets (pre-1.3.0) load fine — use_exp defaults to off and base_alpha defaults to 0.5.

v1.2.0

Added

• Anima block merge — 28 DiT blocks (BLK_00 to BLK_27) in a dedicated mode
• Anima LoRA bake-in — fuses UNet LoRA layers into Anima checkpoints
• Automatic architecture detection at merge time (SDXL vs Anima)
• Bake-in now logs detected architecture to terminal

Notes

• Anima TE keys (Qwen3 text encoder) are skipped during bake-in — the text encoder is a separate file not included in the checkpoint
• Block labels for Anima are positional only (BLK_00–BLK_27) — no documentation exists yet on what each block controls visually

v1.1.0

Terminal logging, stop button, and bug fixes.

What's new

• All merge operations now print progress to the Forge terminal with [NeoMerger] prefix
• tqdm progress bar during checkpoint merges
• Stop button on Merge, Bake-in, and LoRA Merge tabs

Fixed

• Preset dropdown was showing internal bm_ prefix in names
• Preset load crashed with ValueError on old preset format
• Block Similarity Analyzer was showing ~99% similarity on same-family models
• TypeError on Forge base / A1111 (Gradio 3.x)

Installation
Replace scripts/neo_merger.py in your existing installation and reload the UI. No reinstall needed.