Introduction

Forty steps down to six, and the prompt helper folded right in! P-Q-Image-2.1 v0.2, released today by Pruna AI on Hugging Face, is a pair of open-weight LoRA adapters that let Qwen-Image-2.1 generate and edit images in just 6 steps with no CFG. It covers both text-to-image and image editing, and you can run it locally today in ComfyUI (native nodes, no custom packs) or in Diffusers.

The headline trick is the second adapter. Qwen-Image-2.1 normally recommends running a separate prompt-upsampling model first to expand your short prompt. Pruna distilled that upsampler's effect into the LoRA itself, so a short prompt gets the richer look without that extra call.

What shipped

Two new 6-step adapters join the earlier v0.1 pair:

  • 6-steps (p_q_image_2.1_6_steps_v0.2.safetensors): an advanced distillation meant as a drop-in replacement for the base model. Pruna says it beats v0.1 in almost every category and is especially good at editing where you need the input preserved.
  • 6-steps-upsampled (p_q_image_2.1_6_steps_upsampled_v0.2.safetensors): distills the recommended Qwen-Image-2.1 prompt upsampler into the adapter for better visuals from short prompts and a much shorter end-to-end run. Pruna notes it may not exactly match what the real upsampler would write.
  • v0.1 stays up: the 8-step (quality) and 5-step (speed) adapters remain in the repo.
  • Specs: rank 64 across 232 linear targets, loaded at LoRA strength 1.0. Pruna quotes up to 6.3x faster generation, measured on one H100 80GB in BF16.
  • Training coverage: 1K resolution, text-to-image plus single- and multi-image editing with up to 3 reference images. 2K and more references may work but sit outside what it was trained on.

Run it in ComfyUI

Pruna published a separate ComfyUI export with six drag-in workflow files. The adapters load with the native LoraLoaderModelOnly node on top of the Comfy-Org Qwen-Image-2.1 base checkpoint, so you just need a ComfyUI build with native Qwen-Image-2.1 support. Use the base checkpoint, not the Turbo one.

Drop the files here:

  • models/diffusion_models/: qwen_image_2.1_int8_convrot.safetensors or qwen_image_2.1_bf16.safetensors
  • models/loras/: p_q_image_2.1_6_steps_v0.2_comfyui_bf16.safetensors (or the upsampled one)
  • models/text_encoders/: qwen3vl_8b_int8_convrot.safetensors
  • models/vae/: qwen_image_2.1_vae_bf16.safetensors

Then drag in pruna_q_image_2.1_v0.2_6step_t2i.json or pruna_q_image_2.1_v0.2_6step_image_edit.json. To try the upsampled adapter, just switch lora_name in the same workflow.

Settings that matter

  • Sampler euler, cfg 1, no negative prompt, LoRA strength 1.0.
  • Use the v0.2 ManualSigmas schedule and never mix it with the v0.1 ones:

1.0, 0.961054, 0.918949, 0.769302, 0.545554, 0.222292, 0

  • Start at 1024 × 1024 and keep edits to three reference images or fewer.

Run it in Diffusers

The main repo has a Diffusers quickstart on QwenImage21Pipeline (CUDA GPU needed, plus a pinned Diffusers commit listed on the card). The core of it:

pipe.load_lora_weights(
    "PrunaAI/P-Q-Image-2.1",
    weight_name="p_q_image_2.1_6_steps_v0.2.safetensors",
)
SIGMAS = [1.0, 0.925, 0.85, 0.625, 0.375, 0.125]

Pair those raw sigmas with dynamic shifting turned on, true_cfg_scale=1.0, and six inference steps. The card has the full setup, including the scheduler config.

Try this prompt

This is the neon-sign prompt from Pruna's own model card, a nice stress test for text rendering at six steps:

A glowing neon shop sign that reads "QWEN IMAGE 2.1", mounted on a brick wall in a narrow city alley at night. Heavy rain, wet pavement reflecting pink and blue light, shallow depth of field, cinematic photograph.

Licensing

These adapters are a derivative of Qwen-Image-2.1 and ship under the Qwen Research License Agreement, so read its use restrictions before shipping anything commercial. They are also not standalone: you still need the Qwen-Image-2.1 base weights, and Pruna calls this an unfinished release whose weights may change.

Why it matters

Qwen-Image-2.1 already has a crowd of speed-ups, from the Lightning 4-step LoRA to the Turbo 8-step weights. What is new here is folding the prompt upsampler into the adapter. That removes a whole LLM call from the pipeline, which is often the slowest part of a "fast" image model, and it makes lazy one-line prompts look a lot more finished.

Original Source

Conclusion

Six steps, no CFG, and no separate prompt helper make a very tidy local Qwen-Image setup. Grab the upsampled adapter, feed it your laziest prompt, and see how much it fills in for you.

—Aurelia ♡