PRPickrack

AI Image Stylizer (Img2Img)

Re-paint your photo in a new style using Stable Diffusion. Strength slider controls how much to change. Self-hosted GPU.

Stable Diffusion img2img — re-paint your photo in any style. 4 model picker, strength slider, 8 style presets. Self-hosted GPU, free.

AI Image Stylizer uses Stable Diffusion's img2img mode to re-paint your photo in a new style. Upload a photo, type a style prompt (or pick a preset), and the model returns the same scene but rendered in oil paint, watercolor, anime, cyberpunk, Pixar 3D, Ghibli, pencil sketch, or any style you describe. A strength slider controls how much to change — low for subtle stylization, high for bold reimagining.

Use it for: making a profile picture into an oil-painting avatar, turning a vacation photo into a Ghibli landscape, generating album cover variations from a single shoot, prototyping art directions before commissioning an illustrator, social media thumbnail variants, or just creative play with photo → art.

Runs on a self-hosted RTX 3060 GPU. Same Stable Diffusion 1.5 + fine-tunes as the AI-Art QR Generator — DreamShaper 8 (versatile photo + creative), Realistic Vision V6 (hyperreal photo), Anything V5 (anime), or default SD 1.5 (balanced). Pipeline weights shared with QR-art generator → zero additional VRAM cost. ~15-25s per generation.

Key features

  • 4 style models — Default SD 1.5 (balanced), DreamShaper 8 (photo+art versatile), Realistic Vision V6 (hyperreal), Anything V5 (anime). Switch models for radically different aesthetics on the same source image.
  • 8 quick presets — Oil painting, watercolor, anime, cyberpunk, pencil sketch, Pixar 3D, Studio Ghibli, cinematic photo. One click loads the preset prompt.
  • Strength slider — 0.1 (subtle filter, original clearly recognizable) to 0.95 (nearly txt2img with image as loose conditioning). Default 0.65 = balanced — original structure preserved, style boldly applied.
  • Style + content mixing — img2img preserves general composition from your source while injecting style from the prompt. Great for keeping a face / pose / scene while changing the medium.
  • Auto-resize to SD dims — Inputs resized so longest edge = 768px (or your chosen max-side), with both dimensions multiples of 8 (SD UNet requirement). Aspect ratio preserved.
  • Self-hosted GPU — Runs on operator's RTX 3060. ~15-25s/image. Pipeline shared with QR-art generator → no extra VRAM. Free, no signup, no per-generation cost.
  • Browser-side download — Output PNG returned as base64 data URL, downloaded directly. No server storage, no logging of generated content.

How to use

  1. Step 1: Upload photo — Drop or click. JPG / PNG / WebP up to ~6MB. Will be auto-resized to ~768px on longer edge before processing.
  2. Step 2: Pick a style model — Default for balanced output, Realistic Vision for ultra-photo styles, Anything V5 for anime, DreamShaper for everything in between. First non-default model load = +10-30s download.
  3. Step 3: Write a style prompt — Pick a preset (Oil painting, Anime, etc.) or write your own. Best results: describe medium + lighting + mood. E.g., 'Soft watercolor painting, gentle washes, pastel colors, dreamy'.
  4. Step 4: Set strength — 0.5-0.7 = balanced (default 0.65). Lower if you want recognizable original. Higher if you want bold transformation. Experiment with the same prompt at 0.4, 0.6, 0.8.
  5. Step 5: Generate + download — Wait ~15-25s. Side-by-side comparison appears. Download PNG. Try a different model or strength to compare.

When to use

  • Profile picture stylization — turn a phone selfie into an oil-painting portrait or anime avatar
  • Travel photo → art print — vacation shot → Ghibli watercolor landscape, frame and hang
  • Album cover prototyping — band photo → cyberpunk neon variant for cover art exploration
  • Social media thumbnails — generate 4 style variants from one photo for A/B testing engagement
  • Client deliverable mood-board — show 3 art-direction options before commissioning an illustrator
  • Wedding photo stylization — anniversary gift: turn the wedding portrait into a painted artwork
  • Game asset prototyping — photo of a real object → stylized concept art for game design
  • Children's book illustration — sketch your idea, photograph it, stylize as final art mood

Frequently asked questions

How does img2img differ from txt2img?

Txt2img generates from text only (random noise → denoised image guided by prompt). Img2img starts from YOUR image, partially noises it (per strength slider), then denoises guided by the prompt. Result: composition/subject from input photo, style from text prompt. Best of both: precise control over what's in the image (input) AND how it looks (prompt).

What's the strength slider actually doing?

Strength controls noise level. 0.3 = lightly noised input, denoised back with prompt influence — light stylization. 0.7 = heavily noised, mostly regenerated under prompt guidance — bold transformation. 1.0 would be pure txt2img (no input influence). Sweet spot 0.5-0.7 for most cases.

Why does my face look different / weird?

SD 1.5 wasn't specifically trained on faces. Faces often distort slightly, especially at higher strengths (0.7+). For face-preserving stylization, lower strength to 0.4-0.5, OR run the output through a face-restoration model afterward (GFPGAN / CodeFormer — not in Pickrack yet).

Which model for which style?

Photo realism → Realistic Vision V6. Anime/illustration → Anything V5. Mixed media / creative → DreamShaper 8. Unknown / general → Default SD 1.5. Try the same prompt across all 4 models to see massive aesthetic differences — that's part of the fun.

Can I keep the exact composition?

img2img preserves composition reasonably well at lower strengths (0.3-0.5). For STRICT composition preservation (exact pose, exact face position), look into ControlNet-based tools. img2img alone allows the model some freedom — fingers may relocate, hair may change, etc. at higher strengths.

What happens to image dimensions?

Input is resized so longest edge = your chosen max-side (default 768px), with both dimensions multiples of 8 (SD UNet requirement). Aspect ratio preserved. Larger max-side = more detail, more VRAM, more time. 768 is the sweet spot for 12GB VRAM.

Does it work for non-photo inputs (drawings, sketches)?

Yes — img2img works on any RGB input. A pencil sketch + prompt 'detailed oil painting' will return a painted version of your sketch. A 3D render + 'cinematic photo' style → photo-realistic-looking variant. Etc.

How is this different from the AI-Art QR Generator?

AI-Art QR uses ControlNet (QR-Monster) to lock the QR pattern + generate art around it. AI Img2img uses Img2Img mode (no ControlNet) to transform an EXISTING image. Different pipelines, same underlying SD weights (Pickrack shares them in VRAM — zero extra memory cost). Different use cases — QR generation vs photo stylization.

Why are results sometimes random?

Seed is unset by default (random). Same prompt + same input + same settings gives different output each click. This is intentional — multiple shots let you pick the best. Fixed seed is on the roadmap for reproducible iteration.

Can I batch-style multiple images?

Not in v1. One at a time. For 10-50 images: hand-process each (still faster than commissioning an artist). Batch endpoint is on the roadmap.

Related tools

AI-Art QR Generator

Stable Diffusion + ControlNet creates a scannable QR code blended into a custom artwork. Self-hosted GPU.

AI Image Upscaler

Swin2SR transformer 4× upscale with detail enhancement. Better than bicubic for photos. Self-hosted GPU.

Background Remover

Remove image background using AI in your browser. No upload, no signup, no daily limit. Powered by @imgly/background-removal.