Qwen Image 2.1 Turbo · Edit

Turbo means distilled: the same Qwen Image 2.1 editor collapsed from a full multi-step sampling loop into roughly eight steps. The trade is the one distillation always makes - a bit less refinement per render, several times faster and cheaper. Here it works out to about 2.16 credits ($0.0216) per image, under half the price of the plain 2.1 edit, at the same 2K.

That price is the argument. Exploration on a full-fat editor is three cents and a wait per throw; exploration here is a fifth of a cent of dollars and near-instant. Render a pile, judge them, then re-render the winner on the non-turbo node if it needs the extra finish.

What distillation actually costs you

The KB's distillation writeup is blunt about this: fewer steps means less time to refine detail, and every distillation technique loses some quality against the full model. It's not a defect you can prompt your way around. In practice you'll notice it in fine texture, in small text, and in the edges of things that should be sharp - the places where a diffusion model spends its last ten steps.

For a contact-sheet pass, irrelevant. For a final deliverable with a headline in it, go back to Pro or the regular 2.1.

The fields

prompt - same instruction style as the rest of the family: say what changes and, explicitly, what stays. There's no negative prompt on this model, so exclusions go in the sentence. Refer to the sources as "Image 1, Image 2" in the order you connected them.

image_urls - Autogrow, one to ten, ordered. A new socket appears each time you connect something; a batch on a single socket counts as several pictures.

num_outputs - 1 to 6, default 1. Six candidates at 2.16 credits each is about 13 credits, which is the cheap way to actually search a prompt space. Billed per image returned.

resolution - 2K default, "every tier costs the same" - so there's no reason to drop a tier to save money, only to save time.

aspect_ratio - model default by default, meaning the field is left out of the request and the output follows your source images. Leave it, unless you want a reframe.

enable_prompt_expansion - False by default. Turn it on and the service rewrites your brief. enable_thinking - True by default, but per the author's own tooltip it "only works while prompt expansion is on", so out of the box it's decoration. That pairing catches nearly everyone.

seed - fixed. Re-queueing the same graph returns the cached batch instead of charging again.

Outputs: images and task id. No alpha/transparency output on the turbo variants.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart. Or search PopNow in ComfyUI Manager and hit Install. Nothing else installs - zero declared dependencies, because the nodes use the aiohttp, torch, Pillow, NumPy and PyAV ComfyUI already ships. Python 3.10+ and a ComfyUI with the V3 node API.

Key from popnow.ai/api/console/keys into Settings → PopNow, then restart again so the model list comes from your account. It's cheap, not free, and the node list refresh is genuinely a restart thing.

Where it bites

At eight steps the model commits early, so a vague prompt gets a vague image fast. This is the node where prompt hygiene shows most: be specific about what stays, because there's less sampling to average the ambiguity out.

Drift is still the family's weakness - turbo doesn't fix whole-frame re-emission. Chained edits wander, faces included.

PopNow 40201 is an empty balance (2.16 credits still needs a balance), 40004 an unservable combination naming the offending field, 400 a rejected value, 40310 an unverified email. Failed and refused runs are refunded; keep the task id for support. And set Settings → PopNow → Max credits per run - with num_outputs at 6 it's the difference between exploring and spending.