Qwen3-VL-4B for ComfyUI
ComfyUI-format LoRAs for the Qwen3-VL-4B text encoder.
| File | Size | |
|---|---|---|
loras/qwen3vl-4b-heretic_lora_rank1_fp32.safetensors |
2.3 MB | Reproduces DreamFast Heretic on the stock encoder instead of a second checkpoint. Removes refusal behaviour. |
Use
LoRAs into models/loras/, plus the stock bf16 Qwen3-VL-4B text encoder
(qwen3vl_4b_bf16.safetensors) into models/text_encoders/.
CLIPLoader -> LoRA on CLIP (strength 1.0) -> text encoder
- Strength
1.0is the full Heretic model; lower values blend back toward stock behaviour.
How it was made
A plain weight diff, heretic − qwen3vl_4b_bf16, factored per matrix with SVD. Heretic is a
rank-1 abliteration — every changed matrix has one real singular value (~1.8), then a cliff to
bf16 rounding noise (~0.02) — so rank 1 keeps 99.5% of it.
- 62 matrices, language model only:
self_attn.o_proj(layers 4–35) andmlp.down_proj(layers 6–35). Embeddings, norms and the vision tower are untouched.
Keys are ComfyUI's generic text-encoder form
(text_encoders.transformer.model.layers.N.<module>.lora_{down,up}.weight), fp32, no alpha.
Credits
Qwen3-VL-4B-Instruct (Apache-2.0) · DreamFast (Apache-2.0) — all uncensoring behaviour is theirs, this only re-expresses it as a LoRA.