Krita AI Diffusion: Free Local Generative Fill

Short answer: install Krita 5.2+ and the free Krita AI Diffusion plugin (10.4K GitHub stars, GPL-3.0) — you get Photoshop-style generative fill running 100% on your own GPU with no subscription and no credit meter. Minimum hardware is roughly a 6GB-VRAM card (RTX 2060 or better, Apple Silicon M1+, Arc 8GB+, or ROCm AMD per the official README) and 10-50GB of disk. The plugin's installer sets up its own ComfyUI backend for you: download the ZIP, import it via Tools ▸ Scripts, pick a model, select an area on your canvas, click Fill.

That is the whole pitch, and it is why "generative fill free alternative" searches keep landing on this plugin. Adobe's version is genuinely good — but it lives behind a subscription, meters you with generative credits, and rounds every fill through Adobe's cloud. Krita AI Diffusion does the same job on hardware you already own, with models you choose. MakeUseOf ran the head-to-head and titled the result "I ditched Photoshop's generative fill for Krita's free version," calling the output quality a match. This guide is the part those articles skip: the actual setup, verified against the official docs, plus the honest list of where it is worse than Photoshop.

---

What You Get for $0

Answer first: one docker panel inside Krita that does generative fill, object removal, outpainting, ControlNet-guided generation, upscaling to 4K/8K, and a live-painting mode — from a GPL-3.0 plugin that has been actively maintained for 3+ years.

The project is Acly/krita-ai-diffusion — 10.4K GitHub stars as of August 2026, with a steady monthly release cadence (v1.52.1 shipped June 30, 2026; v1.52.0 two days earlier added support for the new Krea 2 model). This is not an abandoned weekend hack; it is one of the most consistently maintained tools in local image generation.

The feature list, per the official README:

• Inpainting — the plugin's own framing is "use selections for generative fill": select, type what you want (or nothing), generate. Also covers object removal.
• Outpainting — extend the canvas and let the model fill the new space to match.
• ControlNet — Scribble, Line art, Canny edge, Pose, Depth, Normals, and Segmentation guidance, mapped onto Krita layers.
• Upscaling — a dedicated workspace for super-resolution up to 4K and 8K.
• Live painting — the model reinterprets your canvas in near-real time as you paint.
• Regions — attach a different text prompt to different areas of the image.

The detail that makes it feel like Photoshop rather than a bolted-on gimmick: results arrive as previews in a generation history, and applying one adds it as a new layer. Your original pixels are never destroyed. It behaves like a native Krita feature because it respects Krita's layer model.

---

Hardware Requirements

Answer first: a 6GB-VRAM GPU is the official recommended minimum, 10GB of disk is the floor (50GB+ with optional packs), and NVIDIA, AMD, Intel Arc, and Apple Silicon are all supported.

From the official README and installation docs (retrieved August 2026):

Table: Acly/krita-ai-diffusion README and docs.interstice.cloud/installation, August 2026.

What the tiers mean in practice: 6GB runs SD1.5 well and SDXL workably; FLUX-class models are flagged by the docs as "relatively large and slow (requires powerful GPU)," so treat them as 12GB+ territory. If you are choosing a card for this workload, our best GPU for image generation guide ranks the realistic options, and the best Stable Diffusion models for 8GB VRAM page shows exactly what a mid-range card can run.

Two platform gotchas straight from the official docs, because they generate a disproportionate share of support threads: on Windows, use the direct Krita download, not the Windows Store or Steam versions; on Linux, use the AppImage rather than Flatpak.

---

Install: The Plugin in 3 Steps

Answer first: download the release ZIP, import it with Tools ▸ Scripts ▸ Import Python Plugin from File, restart Krita, then enable the docker under Settings ▸ Dockers ▸ AI Image Generation.

Verified against the official installation docs:

1. Download the latest plugin ZIP from the GitHub releases page. Do not unzip it.
2. Import in Krita: Tools ▸ Scripts ▸ Import Python Plugin from File…, select the ZIP, approve enabling the plugin, restart Krita.
3. Enable the docker: create or open a document, then Settings ▸ Dockers ▸ AI Image Generation. The panel appears and walks you into backend setup.

That is genuinely all of the Krita-side setup. The real decision comes next.

---

Pick a Backend: Managed Install vs Your Own ComfyUI

Answer first: choose the local managed server unless you already run ComfyUI — the plugin downloads and manages a complete ComfyUI backend for you, which is the closest thing local image AI has to a one-click install.

The docker offers three ways to run, per the official docs:

The managed install is the star. You pick your GPU type (CUDA / ROCm / XPU / MPS), tick which model workloads you want, and it downloads everything — the server, the required ControlNet, inpaint, and CLIP Vision support models, and the checkpoints for your chosen architectures. This is where the 10-50GB of disk goes. When a new plugin version ships, it updates the backend too. If you have ever hand-assembled a ComfyUI install, the difference in friction is dramatic.

The cloud option exists and funds the project, but this site's position is unchanged: the entire point of this plugin is that your canvas never leaves your machine. With a 6GB card you do not need the cloud.

One honest note for tinkerers: under the hood this is standard ComfyUI, and the plugin's Custom Graph workspace even lets you drive your own node graphs from inside Krita. If the node-graph world is new to you, our complete ComfyUI guide covers what the plugin is quietly doing on your behalf.

---

The Generative Fill Workflow

Answer first: make a selection with any Krita selection tool, leave strength at 100% so the action reads "Fill," optionally type a prompt, and hit Generate (Shift+Enter). Results stack up in a history; Apply puts your pick on a new layer.

This is the workflow Photoshop refugees came for, so here it is in detail, verified against the official selections documentation.

Filling and removing. Select the area you want to change — lasso, rectangle, contiguous select, anything. The moment a selection exists, the Generate button becomes a mode dropdown:

• Fill — the general-purpose generative fill, biased toward surrounding colors so results blend.
• Remove Content — object removal: continues the surroundings over the selection. No prompt needed.
• Add Content — for inserting something new; expects a text prompt (or control layers) and allows results that deviate hard from the surrounding colors.
• Expand — outpainting: ignores what is under the selection and extends from the border.
• Replace Background — keeps the foreground, swaps everything behind it (flagged experimental in the docs).

The strength slider is the other half of the tool. At 100%, the action is Fill — pure generation inside the selection. Below 100% it becomes Refine: the model starts from your existing pixels and changes them proportionally less as you lower it. The pattern that works: paint a crude version of what you want, select it, and refine — 30-70% strength is the useful band in our experience. That loop — rough it in by hand, let the model finish it — is the thing Photoshop's one-shot fill cannot do, and it is where a painting app hosting the AI beats a dialog box.

Blending controls. If a fill shows seams, the Diffusion settings expose Selection Feather (falloff as a percentage of selection size), Selection Blend (mask blend size in pixels), and Selection Padding (how much surrounding context is grabbed automatically). You can also switch the context area from Selection Bounds to Entire Image when the model needs to see the whole scene to get lighting right.

Iterating. Every generation lands in the history strip as a preview. Double-click or hit Apply to commit one as a new layer. The docs' own suggested technique is worth stealing: apply two or three candidates as layers and erase between them with a soft brush — you composite the best parts instead of gambling on one perfect roll.

---

Which Base Model to Choose

Answer first: SDXL (RealVisXL or ZavyChroma) if you have 8GB+; SD1.5 on 6GB cards; Illustrious only for anime; Z-Image Turbo as the modern realism pick; FLUX only on a powerful GPU.

During managed setup you tick which model families to install. What the official base-models documentation says about each, condensed:

Table: docs.interstice.cloud/base-models and /models, August 2026.

For generative fill specifically, a practical note: SD1.5 and Illustrious have the richest sets of dedicated inpaint and control models in the plugin's model list, which is part of why fills on those architectures blend so well even on modest hardware. For the FLUX families, check our FLUX VRAM requirements by GPU before downloading 20GB of files your card cannot lift — and if you sit at the 12GB tier, the best FLUX model for 12GB VRAM page maps the quant that fits. For how these families compare outside Krita, see our local image models comparison.

---

Using Your Own ComfyUI

Answer first: point the plugin at any ComfyUI instance — local or on another machine — after installing four required custom node packs and the required models for your architecture.

If you already run ComfyUI, skip the managed install and connect. Per the official ComfyUI setup docs, the plugin requires exactly four custom node extensions:

cd ComfyUI/custom_nodes
git clone https://github.com/Fannovel16/comfyui_controlnet_aux
git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus
git clone https://github.com/Acly/comfyui-inpaint-nodes
git clone https://github.com/Acly/comfyui-tooling-nodes

Two more are optional but worth it: city96/ComfyUI-GGUF (loads quantized .gguf checkpoints — how FLUX fits on smaller cards) and nunchaku-tech/ComfyUI-nunchaku (SVDQuant FLUX variants).

Then install the required models for your architecture from the official model list. The pattern, per the docs: every architecture needs its support models in the usual ComfyUI folders (extra_model_paths.yml is respected). For SD1.5 that means the clip-vision_vit-h.safetensors CLIP Vision file plus the inpaint/control models; FLUX 1 needs Clip-L and T5-XXL text encoders plus flux_vae.safetensors; Z-Image reuses the FLUX VAE. The plugin's connection screen tells you exactly which files it is missing, with download links — let it be your checklist rather than guessing.

The upscale workspace draws on shared upscaler models (NMKD Superscale, OmniSR, HAT variants per the model list). If super-resolution is a big part of your workflow, our local AI image upscaling guide covers those same model families in depth.

---

Honest Limitations

Answer first: this is the best free generative fill you can run, not a free Photoshop. Five things to know before you commit:

1. Krita is a painting app, not a photo editor. No adjustment-layer stack to rival Photoshop's, no Camera Raw. If your day is retouching RAW portraits, Krita hosts the AI beautifully but the surrounding editor is built for illustration first. The fill itself matches Adobe's — the app around it is different.
2. Setup is real. Photoshop's fill needs zero configuration. Here you install a plugin, choose a backend, and download 10-50GB of models. It is a one-time cost and the managed installer does the work, but "one click" it is not.
3. You need a GPU. The 6GB floor excludes a lot of laptops, and CPU mode is officially "very slow." Adobe's cloud does not care what silicon you own; local-first means your hardware is the product.
4. Quality tracks the model you pick, and drift is on you. Firefly is one tuned model with guardrails; here a bad checkpoint choice produces bad fills, and keeping up with better models (the plugin added Krea 2 support in June 2026) is your job. That freedom cuts both ways.
5. It is a volunteer project, not a product. The plugin targets Krita 5.2 or newer and ships on its maintainer's schedule — the latest release at the time of writing is v1.52.1 (June 30, 2026). The project's 3-year track record and monthly releases are reassuring, but this remains a GPL passion project with a Discord, not a company with an SLA.

If you want the local-AI canvas experience with a more photo-editor-shaped UI and can trade away Krita's brush engine, InvokeAI is the other serious artist-facing option — same local models, different philosophy.

---

Verdict

Krita AI Diffusion is the strongest free, local answer to Photoshop's generative fill — full stop. A 10.4K-star, GPL-3.0 plugin with monthly releases, a managed installer that makes ComfyUI invisible, and a selection-fill workflow that third-party testing rates equal to Adobe's, running on a 6GB GPU with nothing leaving your machine.

Set it up like this:

1. Krita 5.2+ (direct download, not Store/Steam builds) + the plugin ZIP via Tools ▸ Scripts.
2. Managed local server; tick SDXL (8GB+ VRAM) or SD1.5 (6GB), add Z-Image Turbo if you have headroom.
3. Learn two habits: Remove Content for cleanups, and rough-paint-then-Refine at 30-70% strength for everything else.

Adobe's version remains more convenient. But convenience is the only column it wins — on price, privacy, resolution limits, credit meters, and model choice, the free local stack takes every row. That trade has a clear winner if you own a GPU.

---

Sources

• Acly/krita-ai-diffusion — official repo: features, hardware requirements, license, star count (retrieved August 2026)
• Plugin releases — v1.52.1 / v1.52.0 (June 2026) release notes
• docs.interstice.cloud — official handbook: installation, selections/generative fill, base models, ComfyUI setup, model list (retrieved August 2026)
• MakeUseOf — "I ditched Photoshop's generative fill for Krita's free version" quality comparison

---

FAQ