Fast AI Image Generation · Speed Add-on

P-Q-Image-2.1: Pictures and Edits in Just 6 Steps

Ask AI ▾

Waiting forty sampling steps for every picture kills the fun of experimenting. This adapter fixes that. Released by PrunaAI, the speed-focused team behind the open-source pruna toolkit, it is a small add-on for Alibaba's open-weights Qwen-Image-2.1 that draws new images and edits existing ones in as few as six steps, with no negative prompt to tune. PrunaAI quotes up to about 6.3× faster generation than the regular 40-step base.

  • Free to start
  • Fast drafts
  • Runs in browser
Start creating images now →
Made by
PrunaAI
Works with
Qwen-Image-2.1 (base weights required)
Latest version
v0.2: 6-steps and 6-steps-upsampled
Does
Text to image and editing with up to 3 references
Speed claim
Up to ~6.3× vs the 40-step base
Open weights
Yes, on Hugging Face

Uncensored Alternative Image Generator

Text to image

Describe the picture you want, or upload a photo and describe the edit.

A turbo button for Qwen-Image-2.1

Think of it as a performance chip rather than a new engine. It is a LoRA, a lightweight file that rides on top of Qwen-Image-2.1 and teaches it to reach a finished picture in far fewer steps. The base model, its text encoder and its VAE stay exactly the same, so you still need the original weights, and you keep everything the base is loved for: sharp text inside images, transparent-background output and instruction-based editing. What changes is the waiting. You set CFG to 1, skip the negative prompt, and get a result in 5, 6 or 8 steps depending on the adapter. PrunaAI measured its headline of up to about 6.3× on a single H100 in BF16 at batch size 1, compared with the usual 40-step run. The project started life as Pruna-Qage-2.1 in September 2026 before moving to its current name.

From v0.1 to v0.2: one step count, two smarter adapters

The first release, v0.1, arrived around September 23, 2026 with a choice between an 8-step adapter, the recommended default, and a 5-step one that is quicker but visibly softer. Version 0.2 replaces that trade-off with a single 6-step schedule pushed further through distillation, then splits it into two flavors. The plain 6-steps adapter holds on to the original photo better when you edit, so faces, logos and product details survive. The 6-steps-upsampled adapter has the official prompt expander baked in, which means a five-word idea comes back as a richly detailed scene without a separate language-model call. The same release brought the first official ComfyUI export with six ready-made workflows; during the v0.1 days, community members had to convert the files themselves.

AdapterStepsStrong pointBest for
v0.1 8-steps8Highest quality of the first releaseSafe default on older setups
v0.1 5-steps5Fastest optionRough drafts where quality can drop
v0.2 6-steps6Keeps the source image faithful when editingOutfit swaps, product edits, portraits
v0.2 6-steps-upsampled6Built-in prompt expansionShort prompts that should look rich

Upsampled vs 6-steps: picking the right adapter

Both v0.2 files share the same size, strength and sampler, so the choice is all about how you like to write. If you write long, precise prompts or you are editing a photo and need the person or product to stay recognizable, load the 6-steps adapter. If you prefer to type a quick idea such as "cyberpunk heroine in a rainy neon alley, cinematic" and let the model invent the lighting, lens and set dressing, the upsampled adapter is the fun one. PrunaAI notes it imitates the official expander closely but not perfectly, so results can differ from running a real prompt-expansion model first.

  • Editing a face, a product or a brand detail: 6-steps.
  • Short, playful prompts and mood boards: 6-steps-upsampled.
  • Maximum speed and quality is secondary: v0.1 5-steps.
  • Most conservative choice on an existing v0.1 workflow: v0.1 8-steps.
  • Load only one adapter at a time and use the sigma schedule that belongs to it.

Training covered 1K images and up to three reference pictures. Start at 1024×1024; 2K output or more references may work but sit outside what the adapters were trained on.

Copy-and-paste prompt templates

These templates follow what works best with the base model: describe the subject first, then setting, light and lens, and put any on-image text in quotes. With the 6-steps adapter use the full template; with the upsampled adapter a short line is enough. They are starting points to adapt, not official samples.

  • Full text to image (6-steps): "[subject], [pose or action], [outfit details], [place and time of day], [light: soft window light / golden-hour backlight], [lens: 35mm half body / low angle], [style: commercial photo / anime], poster title "YOUR TITLE" at the top"
  • Short prompt (upsampled): "cyberpunk heroine on a rainy neon street, cinematic"
  • Outfit swap (18+, fictional or consenting adult): "Keep the face, hairstyle, pose and background unchanged, replace the top with a red string bikini, realistic fabric, lighting matches the original"
  • New background: "Keep the subject and its edges, replace the background with a Maldives beach, add a natural shadow"
  • Three-reference merge: "The person from image 1 wears the outfit from image 2 and stands in the scene from image 3, full body, natural light"
  • Storyboard frame: "Vertical 9:16, [character from image 1], turns and looks back over her shoulder, tense mood, cinematic lighting, shallow depth of field"
  • Product poster: "Product from image 1 on a marble counter, large title "SUMMER DROP" on the left, small line "20% off this week" below, premium soft light"

Writing tip: short, vague prompts look flatter with the plain adapters, so either add detail or switch to the upsampled file.

P-Q-Image-2.1 ComfyUI workflow and settings

PrunaAI publishes a ComfyUI-ready companion repository with BF16 single-file LoRAs and six workflow files: text to image and editing for each version. No custom nodes are needed as long as your ComfyUI build supports Qwen-Image-2.1 natively; the adapter loads through the built-in LoraLoaderModelOnly node at strength 1.0.

Memory: PrunaAI has not published consumer-GPU numbers. The practical lever is the int8 base plus the int8 text encoder, which lowers memory use; quantization and the speed adapter work on different problems and can be combined.

  • Update ComfyUI, then place the base diffusion model (BF16, or the int8 convrot build to save memory), the qwen3vl_8b int8 text encoder, the VAE and the LoRA in their usual folders.
  • Drag the matching official workflow onto the canvas.
  • Sampler euler, CFG 1, empty negative prompt, LoRA strength 1.0.
  • v0.2 ManualSigmas: 1.0, 0.961054, 0.918949, 0.769302, 0.545554, 0.222292, 0.
  • Start at 1024×1024 and only then experiment with larger sizes.
Common mistakeWhat to do instead
Loading the Turbo baseUse the regular base from Comfy-Org/Qwen-Image-2.1; PrunaAI says not to pair it with Turbo
Mixing v0.1 and v0.2 sigmasEach adapter has its own schedule; copy it from that version's workflow
Stacking two Pruna adaptersLoad only one at a time
Jumping straight to 2KBegin at 1K, the size the adapters were trained on

Running it with Diffusers

For Python users, the model card shows a Diffusers path: load QwenImage21Pipeline from Qwen/Qwen-Image-2.1, call load_lora_weights with PrunaAI's adapter repository and the adapter file you want, and use a FlowMatch Euler scheduler with dynamic shifting. Set num_inference_steps to 6, true_cfg_scale to 1 and enable the KV cache; for editing, pass your pictures through the image argument. The Diffusers version uses unshifted sigmas [1.0, 0.925, 0.85, 0.625, 0.375, 0.125] plus dynamic shift. The card pins a specific Diffusers commit and transformers 5.17 or newer, so match those versions to avoid surprises.

How the adapter compares to Lightning, Turbo and int8

People speeding up Qwen-Image-2.1 usually compare three routes. The community Lightning LoRA goes even lower, to 4 steps. PrunaAI's adapter answers with 6 steps, a built-in prompt expander and official editing workflows. The Turbo base is a separate fast checkpoint, and PrunaAI explicitly says not to combine it with this add-on. Int8 or GGUF quantization saves memory rather than steps, so it pairs with either LoRA. No official side-by-side test under identical conditions exists, so the fairest advice is to run your own prompts through both.

RouteStepsWhat you gainWatch out for
PrunaAI v0.2 LoRA6Prompt expansion option, faithful edits, official ComfyUI workflowsTrained at 1K; not for the Turbo base
Lightning LoRA4Fewest stepsNo built-in prompt expansion
Turbo baseFewSpeed without a LoRAIncompatible with the Pruna adapters
Int8 / GGUFUnchangedLower memorySpeeds nothing up on its own

What the community is saying

The upstream repository listed more than 480,000 downloads in the last month, 177 likes, nine discussion threads and use in several Hugging Face Spaces, while the ComfyUI repository collected its first likes within days. ComfyUI Wiki covered the v0.1 launch, and German site Mind-Verse wrote about the release as a way to speed up Qwen-Image-2.1. Before the official ComfyUI export existed, users had already shared unofficial conversions, a sign of how eager local creators were. PrunaAI is candid about trade-offs: v0.1 images fall short of the full 40-step base, short or vague prompts suffer most, outputs may not match the base exactly, and the weights are not final. That honesty is also why v0.2 is getting attention, since it targets exactly those weak spots.

Is there an uncensored P-Q-Image-2.1?

No. The add-on only changes speed; it has no NSFW edition and does not loosen what the base model will draw, which depends on the base weights and any other LoRAs you load. Community uploads labeled "uncensored" Qwen-Image-2.1 have been reported as either the original weights or only a swapped text encoder with no visible change, and the Turbo-based one cannot be used with this adapter anyway. For quick, unrestricted adult edits such as swimsuit or lingerie looks in the browser, Spicy AI offers a similar uncensored image editor.

Swimsuit and bikini swaps

Change an adult's outfit to a bikini or one-piece while the face and pose stay put.

Beach and pool sets

Move a portrait to a sunny beach or a rooftop pool party.

Retro pin-up posters

Polka-dot swimsuits, grain and bold headline text.

  1. Say what must stay the same before what should change.
  2. Name the light and lens: golden-hour backlight, 85mm portrait.
  3. State that every person is an adult (18+).

Upload a photo of an adult or describe a scene and try a bikini, beach or pin-up look online.

Open the image editor →
Before and after outfit edit of an adult woman changed from a summer dress into a red bikini on a sunny beach

Self-hosting the adapter vs generating online on Spicy AI

Local deployment (official weights)Spicy AI online
Content filtersSame as the base model you loadNo content filters
Open weightsYesNot needed — runs in the browser
ModelQwen-Image-2.1 plus the speed adapterSimilar uncensored image editor, hosted online
Price to startYour own GPUFree tier available

FAQ

Is this speed adapter a standalone model?

No. It is a few-step LoRA by PrunaAI that loads on top of Qwen-Image-2.1. The base model, text encoder and VAE stay unchanged, so you need the base weights.

What changed between v0.1 and v0.2?

v0.1 offered 5-step and 8-step adapters. v0.2 settles on 6 steps and splits into a 6-steps adapter that preserves edits better and a 6-steps-upsampled adapter with built-in prompt expansion, alongside official ComfyUI files and workflows.

How much faster is it?

PrunaAI states up to about 6.3× compared with the 40-step base, measured on one H100 in BF16 at batch size 1. It has not published numbers for consumer graphics cards.

Which ComfyUI settings should I use?

Use a ComfyUI build with native Qwen-Image-2.1 support, LoraLoaderModelOnly at strength 1.0, euler, CFG 1, no negative prompt and the ManualSigmas that belong to your adapter. Use the regular base, not Turbo.

Can it edit images with several references?

Yes. Training covered editing with up to three reference images. More references or 2K output may work but are outside the training range.

Was it called Pruna-Qage-2.1 before?

Yes. The v0.1 release was published as PrunaAI/Pruna-Qage-2.1; today it lives under a new name in PrunaAI's Hugging Face account, with a separate repository for ComfyUI files.

Should I choose it or the Lightning 4-step LoRA?

Lightning uses fewer steps. PrunaAI's v0.2 adds a prompt-expansion option and official editing workflows. There is no official same-conditions comparison, so test both with your own prompts.

Is Spicy AI affiliated with PrunaAI or the Qwen team?

No. Spicy AI is an independent platform and is not affiliated with or endorsed by either. All trademarks belong to their respective owners.

Is it free?

You can start free with daily generations, and buy credits when you need more.

Sources

Sign in

Welcome to Spicy AI

or