Krita AI Diffusion

Original Article ↗ Hacker News Discussion ↗

An open‑source Stable Diffusion plugin for the digital painting app Krita is impressing users with Photoshop‑style generative fill, pose editing and in‑painting, all running locally on consumer GPUs or via cloud backends. Commenters compare its capabilities to Adobe’s proprietary AI tools and see it as evidence that free software can quickly match or surpass commercial offerings, though hardware requirements and AMD support remain practical concerns. Much of the debate centers on the ethics and economics of generative AI in art: whether training on existing artwork without explicit consent is acceptable, how these tools affect professional artists’ livelihoods, and if AI will ultimately be just another creative tool or a “labor alienation machine.”

Plugin capabilities & comparisons

  • Many see the Krita AI Diffusion plugin as matching or surpassing Adobe’s generative features (inpainting, pose editing, region-limited edits), undermining any “moat” around Photoshop.
  • Pose editing and iterative character/background workflows impress users, though some note downsides like whole-image regeneration and changing backgrounds unless carefully layered and masked.

FOSS vs proprietary ecosystem

  • Commenters praise Krita and KDE as proof that free/open‑source software can rapidly match proprietary innovation, arguing that marketing, not capability, explains much of proprietary dominance.
  • Krita is distinguished from GIMP as a painting/drawing‑first tool with a more Photoshop‑like workflow; GIMP is seen as more photo‑editing‑oriented and UI‑clunky.

Hardware, performance & platforms

  • The plugin embeds or connects to ComfyUI as backend; it supports CUDA and DirectML, with docs initially implying AMD+Windows only.
  • Several report that ROCm works fine on Linux with external ComfyUI; docs are called misleading rather than the code being incompatible.
  • Performance anecdotes:
    • 4GB VRAM (RTX 3050): ~2 minutes for 2K×2K images; 4K×4K often fails.
    • Older GPUs and even Steam Deck can run SD/LCM on CPU but are very slow (~1 minute per 512×512).
    • High‑end cards (e.g., 4090 + latent consistency models) can reach near‑interactive generation (fractions of a second per low‑step image).
  • Cloud GPUs are supported as a workaround for weak local hardware.
  • macOS support is “untested”; likely blocked more by dev resources and Python ML stack complexity than by Apple Silicon capability.

Krita community stance vs plugin

  • Krita’s official community spaces have explicit restrictions on generative‑AI topics, leading some AI‑tool builders to avoid the platform.
  • Others note developers can’t practically stop third‑party plugins; “a tool is a tool,” regardless of official ideology.

Ethics, consent & training data

  • Many artists are angry about models trained on CC, CC‑BY‑NC, or fully proprietary art without explicit consent; this is framed as theft/piracy and labor alienation.
  • Counter‑arguments:
    • Training is likened to humans learning by studying others’ styles.
    • Expanding IP to ban statistical use of public data is seen by some as a dangerous overreach.
  • Some photographers/coders reject AI specifically because it disregards their chosen licenses or is built on non‑consensual scraping (e.g., parallels to GitHub/Copilot).

Art, creativity & “AI as tool” debate

  • One camp views generative AI as just another medium, akin to photography or photobashing, enabling new forms of art and empowering people without traditional skills.
  • Another camp argues AI “makes creative decisions,” reduces artists to post‑processing “grammar checkers,” and is inherently de‑humanizing and capitalism‑driven.
  • Long sub‑threads debate:
    • Whether using AI undermines the meaning of art as human self‑expression.
    • Comparisons to previous tech shifts (photography, photo retouching, digital tools) and whether those analogies really fit.
    • Whether it’s valid to dismiss artist anger as “moral panic” or “a meme.”

Developers’ analogies (Copilot, coding)

  • Some compare Krita+AI to GitHub Copilot: a productivity tool that removes tedium but demands oversight; others refuse such tools on principle (license concerns, subtle bugs, vendor lock‑in).
  • Several suggest that, as with dev tools, economic pressure may eventually push artists to adopt AI workflows, regardless of personal discomfort.

Access, democratization & careers

  • There’s concern that relying on expensive GPUs could make digital art less democratic, versus the “laptop + cheap tablet + Krita” era.
  • Others counter that:
    • Models are becoming far more efficient and can run on older hardware or rentable cloud GPUs.
    • Most creative careers were always highly competitive; AI may amplify output rather than fundamentally change the tiny share who make a living.
  • Some predict many routine art jobs will be automated, pushing human artists toward higher‑value, more narrative‑ and concept‑driven work; others fear widespread job loss and reduced “artistic fulfillment” as human roles shrink to curation and cleanup.