Qwen-Image 2.1 Fun Acc LoRA: Official 4-Step Distillation

The adapter follows the same recipe Alibaba PAI used for the MiniMax H3 PDD Acc LoRAs in August: take a strong base model, distill it with Parallel Decoding Distillation (PDD, arXiv 2607.26004), and release the extracted LoRA plus the inference code needed to reproduce the reference results. Qwen-Image 2.1 itself was released on September 20, 2026 and is already supported in ComfyUI, so this is an acceleration variant of a model that is only days old.

What ships in the repository

The sigma schedule is part of the release rather than an implementation detail: pdd_config.json pins the four-step schedule to 1.0, 0.9169867, 0.7861579, 0.549491, 0.0, and the exported format is qwenimage21_extracted_prefused_v1. Those numbers are why this is not a plain drop-in LoRA. A generic 4-step sampler with default sigmas does not reproduce the distilled behaviour.

How it compares

The model card publishes three-way comparisons for every example: the teacher at 40 NFE, the PDD LoRA at 4 NFE, and Viggle v0.1 full at 4 NFE. PDD is compared not only against its own teacher but against the other 4-step route for this model, which is the Viggle turbo LoRA released earlier in September.

Teacher: 40 NFE	PDD LoRA: 4 NFE	Viggle v0.1 full: 4 NFE

Text-to-image sample, prompt taken from the PDD paper. All three panels use seed 42.

Editing is covered by the same adapter, including multi-reference edits. The example below re-draws the flag in the reference image, and the third column is the same 4-step route served by the Viggle turbo LoRA.

Reference image	Teacher: 40 NFE	PDD LoRA: 4 NFE	Viggle v0.1 full: 4 NFE

Teacher: 40 NFE	PDD LoRA: 4 NFE	Viggle v0.1 full: 4 NFE

Two-reference edit sample (person and cat), seed 43.

Documented limitations

The model card lists two known gaps against the teacher rather than leaving them to be discovered:

• Dense, small text can degrade noticeably, with distorted character strokes and reduced legibility.
• Some editing outputs look slightly blurrier and darker than the teacher's, with reduced fine-detail clarity.

Availability

There is no official ComfyUI support, and the release is written for the official pipeline:

• Diffusers only: keep the bundled qwenimage21_pdd.py and lora_utils_pdd.py next to the scripts and run python predict_t2i.py (generation) or python predict_t2i_edit.py (editing).
• VideoX-Fun: use a checkout that exposes QwenImage21Pipeline from videox_fun.pipeline and run predict_t2i_videox_fun.py or predict_t2i_edit_videox_fun.py.

For ComfyUI there is only an experimental route so far: Kijai published a Qwen-Image 2.1 PDD branch of the ComfyUI repository, and the LoRA needs conversion before a ComfyUI loader can use it. Because the schedule and the non-LoRA norm_q / norm_k / text_norm parameters are part of the distillation, treat ComfyUI conversions of this adapter as community work rather than a supported path.

The same team's earlier acceleration adapters for MiniMax H3 are the closest reference point for how this family of releases works: there, too, the distilled LoRA shipped with its own inference recipe rather than as a stock LoRA for the base model.