darrellbest/Qwen-Image-2.1-PE-I2I-Heretic

Qwen-Image-2.1-PE-I2I Heretic (abliterated)

Not affiliated with or endorsed by Alibaba / Qwen. A community derivative of `Qwen/Qwen-Image-2.1-PE-I2I`, redistributed under the Qwen Research License (copy included as LICENSE). Non-commercial use only; commercial use needs a separate licence from Qwen.

The image-editing prompt rewriter for Qwen-Image-2.1, a fine-tuned Qwen3.5-VL 9B that turns a short edit instruction plus 1–N input images into a detailed English edit prompt, with its refusal behaviour removed by Heretic directional ablation. bf16, same shapes and parameter count as the source; nothing else was changed.

`system_prompt.txt` is included and required. It defines the output format. It is the unmodified file from the source repo.

Results

The second row is an independent evaluation of the exported weights with Heretic's evaluate_model. Refusals were measured on mlabonne/harmful_behaviors and KL divergence (damage to ordinary behaviour, first-token distributions) on mlabonne/harmless_alpaca, both Heretic's defaults, with Heretic's default system prompt.

With n = 100 the refusal count is noisy (σ ≈ 3 at this rate), so 6 vs 9 vs 11 are not meaningfully different. Trial 998 was chosen from the Pareto front for having the damage level of the widely used T2I Heretic (pottokao/Qwen-Image-2.1-PE-T2I-Heretic, KL 0.036):

Does it still rewrite edits?

Checked through a real editing app's prompt-enhancer path (the official pe_core output contract, enable_thinking, Qwen's sampling settings) on three edit instructions: a plain one ("make the sky a dramatic sunset") and two borderline ones (horror-film fake blood; a prop handgun on a table). Every answer parsed, and the rewrites were as detailed and specific as the original's.

An honest caveat: the original model did not refuse those image edits either. Like the T2I rewriter, it is relaxed about edgy image subjects and near-total in refusing chat-style harmful instructions, which is what the 100/100 baseline measures. The ablation mainly changes the latter.

NVFP4 build for vLLM on Blackwell: darrellbest/Qwen-Image-2.1-PE-I2I-Heretic-NVFP4. Companion for text-to-image: darrellbest/Qwen-Image-2.1-PE-T2I-Heretic.

Reproduce

• —Heretic 3521f86 (main, 2026-09-21), default config except n_trials = 1000 (200, then 800 more on the same study) and export_strategy = "merge". transformers 5.17.0, torch 2.11.0+cu130, one RTX PRO 6000 Blackwell.
• —Trial 998 parameters: direction_scope = global, direction_index = 19.522, attn.o_proj: max_weight 0.977 at 19.045, min_weight 0.935, min_weight_distance 18.014, mlp.down_proj: max_weight 1.456 at 22.104, min_weight 0.831, min_weight_distance 18.551.

Use

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

The family

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.