darrellbest/Qwen-Image-2.1-PE-T2I-Heretic

Qwen-Image-2.1-PE-T2I Heretic (abliterated)

Not affiliated with or endorsed by Alibaba / Qwen. A community derivative of Qwen/Qwen-Image-2.1-PE-T2I, redistributed under the Qwen Research License (copy included as LICENSE). Non-commercial use only; commercial use needs a separate licence from Qwen.

The text-to-image prompt rewriter for Qwen-Image-2.1, a fine-tuned Qwen3.5-VL 9B that turns a short request into a detailed English prompt plus a recommended aspect ratio, with its refusal behaviour reduced by Heretic directional ablation. bf16, same shapes and parameter count as the source.

system_prompt.txt is included and required (the unmodified file from the source repo).

Companion: darrellbest/Qwen-Image-2.1-PE-I2I-Heretic for image edits.

Results

	Refusals	KL divergence
Original	98/100	0 (by definition)
This model, re-measured after export	31/100	0.0130
This model's trial during the search	24/100	0.0131

Measured by Heretic on mlabonne/harmful_behaviors (refusals) and mlabonne/harmless_alpaca (KL divergence, the damage to ordinary behaviour), with Heretic's default system prompt. The first row is an independent evaluation of the exported weights (evaluate_model); the search's own figure for the same trial was 24/100. The KL matches exactly; the refusal gap is within about 1.6σ at n = 100. For reference, the same evaluation of pottokao/Qwen-Image-2.1-PE-T2I-Heretic measured 5/100 at KL 0.0349 (its card: 3/100, 0.036), a stronger ablation with more side effects.

This is a light-touch ablation. Across 1,000 trials the search's Pareto front bottomed out at 23/100 refusals, so this model keeps roughly a quarter to a third of the original's chat-style refusals, while changing its ordinary behaviour less than most ablations (KL 0.013; for comparison the I2I companion is 9/100 at KL 0.037). If you need fewer refusals, pick a stronger ablation. Pareto front of the search:

Refusals	KL
23/100	0.0194	trial 228
24/100	0.0131	trial 225 (this model)
30/100	0.0089	trial 223
36/100	0.0085	trial 245
46/100	0.0044	trial 231

With n = 100 the refusal count is noisy (σ ≈ 4 here), so 23 vs 24 is no difference; 225 was chosen for its lower KL.

Does it still rewrite prompts?

Checked through a real image app's prompt-enhancer path (the official pe_core output contract, enable_thinking, Qwen's sampling settings) on a plain request and two borderline ones (a horror poster with a bloodied survivor; a soldier aiming a rifle). Every answer parsed, with a detailed prompt and an aspect ratio. As with the other Qwen-Image-2.1 rewriters, the original is already relaxed about edgy image subjects; the ablation mainly changes its refusals of chat-style harmful instructions, which is what the 98/100 baseline measures.

NVFP4 build for vLLM on Blackwell: darrellbest/Qwen-Image-2.1-PE-T2I-Heretic-NVFP4.

Reproduce

• Heretic 3521f86 (main, 2026-09-21), default config except n_trials = 1000 and export_strategy = "merge". transformers 5.17.0, torch 2.11.0+cu130, one RTX PRO 6000 Blackwell.
• Trial 225 parameters: direction_scope = global, direction_index = 17.498, attn.o_proj: max_weight 1.49 at 21.632, min_weight 0.958, min_weight_distance 16.843, mlp.down_proj: max_weight 1.452 at 21.259, min_weight 0.001, min_weight_distance 10.638.

Use

Exactly like the original: load with AutoModelForImageTextToText / AutoProcessor and use system_prompt.txt as the system prompt.

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.