5 Best Uncensored AI Image-to-Image Generators in 2026

Key takeaways

• Running Qwen-Image-Edit in ComfyUI on your own graphics card is the only image-to-image setup with no platform filter. It's free and released under Apache 2.0, but you'll want a 16GB card, and the base model needs community LoRAs for nudity.

• Mage is the best hosted pick for most people. Inpainting and Refine start on its $10 Basic plan, and its Characters and References tools keep a face, outfit or location consistent from a single uploaded image. NSFW content is supported.

• No graphics card? ThinkDiffusion rents you a cloud machine running ComfyUI or Forge in the browser from $0.99 an hour, and its terms carry no clause on adult content, only a ban on unlawful use.

• Flux Kontext and FLUX.2 are the best-known open image editors, but Black Forest Labs filtered NSFW material out of their training data, and their licenses require content filters or manual review. That makes them strong general editors with little to offer here.

What image-to-image means in 2026

The label covers three different jobs. Classic img2img takes an uploaded picture and a prompt, and a denoise (or strength) setting decides how far the result moves from the original: low values keep the composition, and a value of 1.0 all but ignores it. Inpainting is the same idea limited to an area you mask.

The second job is instruction editing. Models such as Qwen-Image-Edit and Flux Kontext take an image and a sentence ("make the jacket red leather, keep everything else") and change only what you asked for. The third is reference generation, where you upload a face, outfit or location and the model builds a new image around it.

Uncensored tools differ most on the second and third jobs, so that's where this ranking spends its time. For text-to-image realism, see our ranking of the best photorealistic NSFW AI image generators.

How we ranked

We weighed each tool on how much it allows on uploaded images, the quality and control of its edits (masking, references and instruction editing), what a month of regular use costs, how you pay, and how it treats real people's likenesses.

The five tools compared

1. Self-hosted ComfyUI + Qwen-Image-Edit - Best for full control

If you want nothing between your upload and the output, you'll have to run the model yourself, and in 2026 that mostly means ComfyUI.

What works: ComfyUI is free, released under GPL-3.0 and updated roughly every week; version 0.39.0 shipped on October 5, 2026. It handles img2img with a denoise setting, inpainting and the newer instruction editors natively, and its docs include a ready-made image-to-image workflow. The editor we'd start with is Qwen-Image-Edit-2511, a 20B model from Alibaba's Qwen team under Apache 2.0, which accepts several input images at once and holds characters steady between edits. Its 4-bit quantized versions run about 12 to 13GB, so a 16GB card is the practical floor. For classic img2img and inpainting, SDXL checkpoints such as Pony and Illustrious still have some of the deepest libraries of community fine-tunes and LoRAs (Low-Rank Adaptations, small add-on models).

What doesn't: Qwen-Image-Edit doesn't render nudity well on its own, so users add community LoRAs or merged checkpoints. Flux Kontext dev and FLUX.2 dev, the best-known open editing models, had NSFW content filtered from their training data, and their non-commercial licenses require content filters or manual review of outputs. Node graphs take an afternoon to learn. Nobody moderates what you make, so the legal responsibility is yours alone.

Best for: Technical users with a 16GB or larger graphics card who want complete control.

2. Mage - Best for people without access to a graphics card

Mage is a browser-based image and video platform that puts most of the image-to-image toolkit behind one subscription. There's nothing to install, and it's uncensored.

What works: Inpainting (mask an area and describe the change) and Refine (a whole-image pass from a description) start on the $10 Basic plan, and editing is unlimited on every membership, on the models each tier includes. Separate apps handle outfit swaps, character swaps, face swaps and relighting from an uploaded image, and each of them does its job effectively. There's also an agent that can streamline the generation workflow.

Characters keeps one face consistent from a single reference photo, and References does the same for outfits, objects, poses and locations, which deals with identity drift, the problem of a face changing from one edit to the next. Mage's own Mango 3 model, unlimited from the $30 Pro plan, takes up to 10 reference images. Open models such as Flux 2, Krea 2 and SDXL sit alongside it with LoRA support, and Enhance upscales to 2K on Pro and 4K on Pro Plus. Generations on a paid membership are private by default.

What doesn't: Some hosted models on the platform, including Nano Banana 2 and GPT Image 2, keep strict safety filters of their own.

Best for: Anyone who wants masking, swaps and consistent characters in a browser, paid for with a normal card.

3. Forge Neo - Best for local editing without nodes

If ComfyUI's node graphs put you off, Forge Neo keeps the familiar Stable Diffusion WebUI layout, with img2img and inpaint tabs and a denoising slider.

What works: Forge Neo is an actively maintained fork of Forge, free under AGPL-3.0, and it runs newer models including Flux Kontext, Qwen-Image-Edit, Z-Image, Krea 2 and Chroma alongside SDXL. It uses the same community checkpoints and LoRAs as ComfyUI. InvokeAI is the other no-nodes option, an Apache 2.0 app built around a canvas for inpainting and outpainting.

What doesn't: The original Forge hasn't been updated since June 2025, so use the fork. InvokeAI's founding team joined Adobe and its hosted service shut down, though the open-source app is still updated, with a 7.0 alpha out in October 2026. Like ComfyUI, both need your own graphics card and leave moderation to you.

Best for: Local users who want img2img and inpainting in a familiar tabbed interface.

4. PixAI - Best for anime edits

PixAI is an anime-focused generator that allows NSFW content as long as you tag it.

What works: Members get Flow Edit and PixAI Reference Pro, which edit uploaded images by chat instruction and keep a character, pose and outfit consistent. An inpaint and outpaint editor sits alongside them. As a member you'll also get three to 10 monthly LoRA training slots depending on plan, and unused credits roll over. Every user gets daily free credits.

What doesn't: It's built for anime and manga, so it's the wrong tool for photos. The chat-based editors, Flow Edit and Reference Pro, are members-only.

Best for: Anime artists who want to edit and restyle characters.

5. ThinkDiffusion - Best for no graphics card

ThinkDiffusion rents you a cloud machine that runs ComfyUI, Forge or Automatic1111 in your browser, so you get the open-model setup from entry 1 without owning the hardware.

What works: You can load any models and extensions, including Qwen-Image-Edit and community LoRAs, and ThinkDiffusion advertises unrestricted models and extensions on its homepage. There's no subscription on the Hobby plan, which includes 15 free minutes and then charges $0.99 an hour for a 16GB machine, $1.75 for 24GB and $2.50 for 48GB. TD-Pro ($29.99 a month, or $19.99 a month billed yearly) cuts those rates to $0.79, $1.49 and $1.99 and adds 200GB of storage. Its terms require users to be 18 or older and ban unlawful use, with no separate clause on adult content.

What doesn't: You're paying for every hour the machine runs: at $1.75 an hour, 18 hours on a 24GB machine costs more than Mage's $30 Pro plan. You'll still find, download and wire up the models yourself, and nobody moderates what you make, so the legal responsibility is yours.

Best for: Users without a graphics card who want open models and no platform filter.

Didn't make the cut

RunPod and RunComfy: they're popular ways to rent ComfyUI, but their terms ban pornographic content.

Grok Imagine: in January 2026, after investigations in the UK and California, X said it had stopped the Grok account on X from editing real people into revealing clothing, though early tests found the standalone Grok app still allowed it.

Tensor.Art: went safe-for-work only on November 27, 2025, and moved adult content to a separate site, TensorHub, without its workflow builder.

How to choose an uncensored image-to-image generator

Decide first whether you have a graphics card with 16GB or more, and whether you need explicit output or only permissive edits.

• No filter, and a strong GPU: ComfyUI with Qwen-Image-Edit.

• Local, but sliders instead of nodes: Forge Neo.

• Masking, swaps and consistent characters in a browser: Mage.

• Open models with no graphics card: ThinkDiffusion.

Frequently asked questions

Is Flux Kontext uncensored?

No. Black Forest Labs filtered NSFW content from Kontext's training data, ships an integrity checker with it, and its license requires content filters or manual review. FLUX.2 dev carries the same requirement. Both are strong editors for everything else.

What's the best open-source model for uncensored image editing?

Qwen-Image-Edit-2511. It's released under Apache 2.0 and accepts several input images at once. It doesn't render nudity well on its own, so users add community LoRAs. For classic img2img, SDXL fine-tunes such as Pony and Illustrious have years of community LoRAs behind them.

What GPU do I need for local image-to-image?

A 16GB card is the practical floor for a quantized Qwen-Image-Edit, whose 4-bit files run about 12 to 13GB. A 4-bit Flux Kontext dev file is about 7GB. FLUX.2 dev is far heavier: it has 32B parameters, and without offloading it won't run on less than 80GB of video memory.

What does denoise strength do in img2img?

It sets how far the output moves from your uploaded image. Low values keep the composition and change details, and a value of 1.0 all but ignores the original.

Is it legal to edit photos of real people with an uncensored AI tool?

Not for sexual edits without the person's consent. In the US, publishing nonconsensual intimate images is a federal crime under the TAKE IT DOWN Act, and platforms must remove them within 48 hours of a valid request. In England and Wales, creating a sexual deepfake of an adult without consent has been an offense since February 6, 2026, even if it's never shared. Mage prohibits deepfakes, and PixAI bans content that invades privacy or publicity rights.