raspie/gemma4-tipo-lora-pilot-001

Gemma 4 TIPO LoRA — pilot 001 (experimental)

Small language-only LoRA, trained for Danbooru-style prompt/tag expansion. This is an experimental candidate, not the default model, not a full checkpoint and not a vision adapter. It is not trained for the newer paired natural-language/tag task. Tag-specialized output can interfere with prose requests.

한국어 / 사용 방법

원본 모델을 재배포하지 않고 LoRA만 제공합니다. 별도 LoRA 로더 노드는 사용하지 않습니다.

1. RS Seasonal Prompt Generator를 업데이트합니다.
2. 아래 정확한 베이스를 ComfyUI 네이티브 safetensors로 준비하여 Load CLIP에 연결합니다. GGUF/mmproj는 이 경로에 사용할 수 없습니다.
3. Gemma TIPO Native 또는 Gemma Prompt Expander Native 안의 **tipo_lora = pilot-001**를 선택합니다.
4. 실행하면 노드가 작은 LoRA와 설정만 다운로드하고, 기존 CLIP의 복제본에 적용합니다. 이후에는 로컬 파일을 재사용합니다. 기본값 none은 LoRA를 적용하거나 다운로드하지 않습니다.

공유 자료에는 개인 이미지·VRS 경험·사전·학습 답안이 들어 있지 않습니다. 이름 보존, 한국어 제외 지시, 이미지 해석의 오류가 남아 있습니다. 새 이중 출력/VRS 기능의 성능 향상을 입증한 모델은 아닙니다.

Exact base and native preparation

• Base: p-e-w/gemma-4-E2B-it-heretic-ara, revision c9a1d4c031981f14d86eeb0c7d87de7fafd34513.
• This is the Heretic ARA derivative, not the unmodified Google instruction checkpoint.
• Upstream lineage: google/gemma-4-E2B-it.
• HF base weight SHA-256: 15605c16e479a8dc53928dd4b521a4ead9fe66222dc14d1ce392384e999a544b.
• Native conversion preserves every tensor byte, precision and tokenizer. Do not apply this adapter on an already merged pilot checkpoint; that would apply the delta twice.
• Other E2B checkpoints can have identical shapes; the node cannot identify source weights from shape alone. Use the exact base above.

Download that revision (using the website, or hf download p-e-w/gemma-4-E2B-it-heretic-ara --revision c9a1d4c031981f14d86eeb0c7d87de7fafd34513 --local-dir /path/to/base) then run the repository's existing converter:

python scripts/prepare_gemma4_native.py --source /path/to/base --output /path/to/ComfyUI/models/text_encoders/gemma4-e2b-native-bf16.safetensors

In Load CLIP, use type=stable_diffusion, device=default. The TIPO node handles its adapter internally using ComfyUI ModelPatcher. No PEFT, Transformers, bitsandbytes or llama.cpp is needed for native inference; PEFT format is preserved for other tools. No additional loader node or merged model download is needed. The full base is still required locally; an adapter does not replace it.

Offline: place adapter_config.json and adapter_model.safetensors in ComfyUI/models/loras/rs_gemma_tipo/pilot-001/ before running; the node checks published SHA-256 hashes. Other local files are not silently overwritten.

See native usage and installation and two output formats.

Training and evaluation

• Independently trained from the same base (pilot 002 does not continue pilot 001).
• 320 text-only supervised training examples, one pass, rank 8 / alpha 16, 11,857,920 trainable parameters.
• 190 language attention/MLP weight targets. Vision, projector, audio, PLE and embeddings were untouched.
• Data focuses on 2D animation/illustration and stylized 3D prompt tasks; metadata is not verified visual ground truth.
• No final holdout was used. Reported metrics use 100 development cases and the same posture-conflict audit for all models.

Model	Raw condition passes / 100	After tag cleanup / 100	Specified names preserved / 245
Base	15	18	226
Pilot 001	42	50	244
Pilot 002	50	54	236

These are automatic tag-condition checks on the separately merged BF16 models, not human quality, character-identification or general language scores. Native runtime patching uses ComfyUI's weight/casting lifecycle; exact token-for-token parity with merged BF16 is not promised. Pilot 002 improved some expansion/edit checks but regressed on protected names and exclusions. Neither candidate is selected as the default.

training_summary.json contains sanitized provenance. Tensor bytes are unchanged from the original adapter. The release config replaces the private filesystem path with the public base and restores exact language-only target names from the training run.

License

Apache-2.0; see LICENSE and NOTICE. Upstream base/lineage model cards are linked above. This repository contains only adapter weights and documentation, not the base model or any personal source data.