Qwen-Image-2.1 PE Heretic ComfyUI single-file builds
The refusal-ablated Qwen-Image-2.1 prompt rewriters, repacked as the single .safetensors files ComfyUI loads.
Same weights as the safetensors releases (T2I,
I2I), merged from their shards: 760 tensors,
identical names and values, nothing renamed or quantized.
Not affiliated with or endorsed by Alibaba / Qwen. Derived from Qwen/Qwen-Image-2.1-PE-T2I and -PE-I2I under the Qwen Research License (copy included). Non-commercial use only.
These rewrite your prompt; they do not generate images. You still need Qwen-Image-2.1 itself for that.
Files
| File | Size | What it rewrites |
|---|---|---|
qwen_image_2.1_pe_t2i_heretic_bf16.safetensors | 17.5 GB | a short request → a detailed prompt + aspect ratio |
qwen_image_2.1_pe_i2i_heretic_bf16.safetensors | 17.5 GB | an edit instruction, reading up to 10 input images |
qwen_image_2.1_pe_t2i_heretic_fp8_e4m3fn.safetensors | 10.7 GB | same, 8-bit |
qwen_image_2.1_pe_i2i_heretic_fp8_e4m3fn.safetensors | 10.7 GB | same, 8-bit |
FP8 builds: the 358 linear weight matrices are cast to float8_e4m3fn; embeddings, lm_head, the patch/position
embeddings and all 1-D tensors stay bf16. Mean error on the cast weights is 4.8%. Checked before publishing: the FP8
weights still produce a valid rewrite with the same aspect ratio (398 vs 472 words on the same prompt). This is a
plain cast, not a scaled or calibrated FP8, so use bf16 if you see drift.
system_prompt_t2i.txt and system_prompt_i2i.txt are included. Each task has its own and they are not
interchangeable; the node ships the stock ones, which apply unchanged here.
Use in ComfyUI
- Put the file(s) in
ComfyUI/models/text_encoders/. - Install the prompt-enhancer node: benjiyaya/ComfyUI-Qwen-Image-2.1-Prompt-Enhancer.
- Pick the file in the node (it loads through ComfyUI's
CLIPLoader, typeqwen_image). The I2I node takesimage_1…image_10, referenced in the instruction as<image1>,<image2>, …
These are drop-in replacements for the stock single-file builds (e.g. HarleyWang/Qwen-Image-2.1-PE-ComfyUI); the only difference is the ablation.
Smaller than FP8? The GGUF builds (Q8_0 ≈ 8.9 GB, Q4_K_M ≈ 5.2 GB) work with llama.cpp/Ollama-based
LLM nodes rather than CLIPLoader; see the family table. There is no 4-bit format that ComfyUI's CLIPLoader
reads, so 4-bit means GGUF (LLM nodes) or the NVFP4 builds (vLLM). An int8-ConvRot build is not published here.
Ablation
Refusal rates and the damage they cost (KL divergence) are on the parent cards: T2I · I2I.
The family
| Model | Format | Use it with |
|---|---|---|
| PE-T2I-Heretic | bf16 safetensors | transformers / diffusers |
| PE-T2I-Heretic-GGUF | GGUF BF16 / Q8_0 / Q4_K_M | llama.cpp, Ollama |
| PE-T2I-Heretic-NVFP4 | NVFP4 (compressed-tensors) | vLLM on Blackwell |
| PE-I2I-Heretic | bf16 safetensors | transformers / diffusers |
| PE-I2I-Heretic-GGUF | GGUF BF16 / Q8_0 / Q4_K_M (+ mmproj) | llama.cpp, Ollama |
| PE-I2I-Heretic-NVFP4 | NVFP4 (compressed-tensors) | vLLM on Blackwell |
T2I rewrites a short request into a detailed prompt for new images; I2I rewrites an edit instruction, reading the image being edited. Both are prompt rewriters for Qwen-Image-2.1, not image generators.