Everything you need to make great pixel art.

1. Pick a model.On the Home page, choose from the Model panel on the left. RD Pro is the highest quality and supports reference images, RD Plus is a strong all-rounder, RD Fast is the quickest and cheapest, and RD Mini specializes in tiny, low-resolution art. The Advanced Animations, Animations, and Tilesets tabs live in the same panel.
2. Pick an art style.Each model has its own set of styles in the Art Styles panel — Default, Painterly, Isometric, Top Down, and so on. Click a style to see its description, and use View Examples to preview what it produces.
3. Describe what you want.Write a prompt about the subject — the style is already handled by your model and style choice. Be specific about details, colors, and setting. Stuck? The sparkle button expands a short idea into a detailed prompt.
4. Set size and count, then Generate.Pick your canvas size in pixels and how many images to make. The exact price always shows on the Generate button before you commit — no surprises.
5. Refine your results.Results appear on the canvas and in your History on the right. Use the editing tools next to the canvas to inpaint, expand, remove backgrounds, reduce colors, or make textures tile. Pin favorites, download as PNG, or drag any result back in as an input image.
6. Go further.Build your own custom style from reference images in My Styles, or grab an API key from Developer Tools to generate images from your own code and tools.

The top-right corner

• Balance — your account balance in USD. Every generation's cost is deducted from it. Click it to top up.
• Account menu — payment methods, invoices, avatar creation, account settings, and sign out.

1Retro DiffusionHome · Guide · Gallery · Developer Tools · Affiliate · My Styles$12.345You

2 Model

RD FastRD PlusRD ProRD MiniAnimationsAdvanced Animations

TilesetsMy Styles

Professional grade pixel art model

3 Art Styles

DefaultPainterlySimpleFantasyHorrorSci-fiIsometric…

Clean modern pixel art style model…View Examples

4 Prompt

A raven standing on a branch with appalachian forests…

5 Reference Images (0/9)

+

6Image Count − 1 +

Generate · $0.18

8

7

9

10 Pinned

fireball orbdiamond swordexplosion vfx

11 History

Upload ImageA small alchemist's cottage…A floating sky island village…A riverside village with a market…

Left: the creation panel (2–6)

Work top to bottom: pick a Model (or the Animations / Tilesets / My Styles tab), pick an Art Style, write your Prompt, adjust Settings (size, seed, tiling, background removal), set the Image Count, and press Generate. The button shows the exact price of the batch before you click.

Generations run through a queue — you can keep queueing more batches while earlier ones are still running ("Add to Queue"), and a status line shows how many are pending and running.

Middle: the canvas (7–9)

The canvas shows your selected generation. Drag to pan, scroll or pinch to zoom, double-click to zoom in. The toolbar above the image gives you, from left to right:

• Reset Position / Zoom In / Zoom Out — view controls.
• Download Image — saves an upscaled PNG (or GIF for animations). Download at original size saves the raw pixel dimensions.
• Copy Image — straight to your clipboard.
• Apply Settings (left arrow) — loads that generation's prompt, style, size, and seed back into the creation panel. Great for iterating.
• Copy Settings Link (link) — copies a browser-only URL fragment that loads the selected image's prompt, style, size, seed, and visible options for another signed-in user. Generations that depend on uploaded or reference images cannot be copied because those assets are not embedded.
• Pin (bookmark) — keeps the image in the Pinned panel for quick access.
• Download Batch — zips all images from a multi-image generation.
• Delete (trash) — removes the generation, with a confirmation.
• Info menu (⋮) — shows the full generation details (prompt, style, seed, size) and a Show API Code button that produces ready-to-run code for recreating the generation through the API.

Along the side of the canvas is the editing tool rail — inpainting, outpainting, background removal, and more. Those are covered in Editing Tools

Right: history and pins (10–11)

• Every generation appears in History, newest first. Click one to load it onto the canvas.
• The bookmark icon on a history item pins it to the Pinned panel above.
• Drag and drop: history and pinned thumbnails can be dragged straight into the input image or reference image slots in the creation panel.
• Upload Image (top of History) — bring your own images into the workflow to edit, pixelate, or use as references.

How it works

1. Ask.Type a request like "a knight sprite, then make it walk". Paste, drag, or attach images when you want it to work from your art.
2. Review the plan.Every step shows its model, size, and price, plus the total. Nothing is charged until you approve.
3. Approve, deny, or ask for changes.Replying with feedback gets you a revised plan.
4. Watch it run.Steps go through the same queue as your own generations, and results appear in the chat, your history, and the canvas. The assistant checks every result and tells you honestly when something came out wrong.

What it can do

• Multi-step work: a base sprite and then its animation, a set of matching items, variants of one image (seasons, palettes, outfits), and edits chained onto earlier results.
• Editing: the canvas edit tools, plus pulling single sprites out of an uploaded sheet.
• 3D, when you ask for it: build a Low-Poly model, revise it, animate it, or change one of its existing animations. Say "3D model" or "low poly"; otherwise it makes 2D art.
• The model on your canvas: when a 3D model is open, "animate this" or "make it blue" works on it directly. 3D results in the chat have an Open in 3D view button.

What it costs

Chatting is free. Approved plans bill at the normal price of each generation, edit, or 3D job, and failed steps are refunded automatically. After 20 messages without generating anything, chat pauses until something is generated. The same assistant is available to programs through the assistant API ($0.01 per message).

What works

• Subject first, then details: "A raven standing on a branch, with Appalachian forests and mountains in the background."
• Name concrete visual features: materials, colors, lighting, mood. "A hammer made of obsidian with a gold handle and a red gemstone."
• Match the style's intent: asset styles (Game Asset, Top Down Item, MC Item) work best when you describe a single object and mention a plain background; scene styles (Environments, Top Down Map) want a full scene description.
• Use the placeholder as a template. Every style shows an example prompt in the prompt box — it's written to match how that style was built.

What to avoid

• Don't put sizes in the prompt ("32x32", "16 bit") — use the width and height settings instead. The prompt box will warn you if it spots this.
• Don't describe motion in image styles — for walking, attacking, or effects, switch to the Animations tab.
• Don't stack contradictions — "minimalist, extremely detailed" pulls the model in two directions.

Handy prompt tools

• Expand Prompt (sparkles button) — turns a short idea into a detailed prompt, typed out live so you can watch and edit it after.
• Clear prompt (trash button) — empties the box; Ctrl+Z brings the old text back.
• Seeds — reuse the same seed with small prompt tweaks to iterate on a composition you like. Find a generation's seed in the canvas info menu (⋮).

RD ProHighest quality

The professional-grade model. Extremely detailed prompt following, the cleanest pixel work, and support for up to 9 reference images to guide style and content. Includes the Edit and Pixelate styles for transforming existing images. Best for hero assets, portfolio pieces, and matching an existing art style.

96–256 pxup to 4 per batchhighest cost

RD PlusAll-rounder

The advanced general-purpose model with the largest style library — from Watercolor and Cartoon to Isometric, Top Down maps, Environments, and Minecraft textures. Great quality at moderate cost, with large batches for exploring variations.

64–384 px (styles vary)up to 16 per batchmoderate cost

RD FastSpeed & volume

The standard model, tuned for speed and price. Ideal for brainstorming, iterating on prompts, and generating lots of options cheaply before refining the winners with a bigger model.

64–384 px (styles vary)up to 15–16 per batchlowest cost

RD MiniTiny art

Optimized for genuinely low-resolution art — 16 to 192 pixels. Icons, items, Minecraft-style textures, and small assets where every pixel counts. Includes fast variants of its styles for quick drafts.

16–192 pxup to 16 per batchlow cost

Each card above ends with the same prompt and seed run on that model, so you can compare them directly.

Every scale, one site

From a 16×16 inventory item to a 384×384 scene — shown here at actual pixel size:

Specialized tabs

Advanced Animations in action — a still frame in, a living animation out:

→

→

→

And from the regular Animations tab — full sprite formats from nothing but a prompt:

Making a model

1. Pick Low Poly 3D in the model selector, then a style: Detailed (free-form shapes) or Blocky (boxes only, like a Minecraft model; boxes can be tilted for slopes like a windshield or roof, and it is best for Blockbench and Minecraft exports). Their Lite versions make simpler models for $0.10 at any size, great for props and drafts; editing or animating a Lite model is $0.08.
2. Describe the object (up to 250 characters), add up to 4 references, or both. Images guide the look (several views of the same object work best). Drag a 3D model from your history into References to use that exact model, textures and all: it is placed where your prompt puts it (a table in a tavern) or adapted to fit.
3. Optionally add a Palette (the same palette input as for images): the model is built with only its colors, and revisions and animations keep them.
4. In Settings, choose a Size, the model's largest dimension in texture pixels. Auto reads your prompt (a coin is small, a castle is big); image-only requests use 64³. Simple requests ("a crate", "remove the text", "spin slowly") are built in a quicker mode automatically; anything detailed gets the full treatment. The price is the same either way.
5. Generate. 3D models take a few minutes (large sizes can take 10 or more). They wait in the queue like any generation, and the finished model opens in the canvas.

The 3D canvas

• Drag to orbit, right-drag to pan, and scroll to zoom. The toolbar has reset, turntable, perspective/orthographic, X / Y / Z view snaps (click again for the opposite side), bones,wireframe, and copy (a square image of the current view, on your clipboard for sharing).
• Switch between the 3D view and the texture atlas, the flat image every face is painted from.
• Export for… Godot, Unity, Unreal, the web, or Blender (.glb), Blockbench (.bbmodel), other 3D apps (.obj), a Minecraft Java resource pack, the atlas, a turntable sprite sheet, or any animation as a GIF, sprite sheet, or high-res MP4 video. The .glb and Blockbench files carry every animation, ready to play in the engine.
• Re-rig (in the tool sidebar, on rigged versions) rebuilds the skeleton: move a pivot, add bones, or leave the prompt empty for a fresh rig. Animations whose bones stay keep working.
• Split (in the tool sidebar) breaks the model into separate models: a desk with things on it becomes the desk, the candle, each book, and so on, each its own history item facing the front, with its own textures and every animation it was in. Leave the prompt empty to split off every object, or say how ("only the books", "each drawer too").
• Color Reducer, Palette Converter and Color Style Transfer (in the tool sidebar, the same tools as on the image canvas) recolor the model's textures: reduce the number of colors, convert to a palette, or take another image's color style. The result is a new version, and later revisions and animations keep the colors.

Revising and animating

• Revise (left toolbar): describe what should change. The result is saved as a new version, and you can switch between versions at the bottom of the canvas.
• Animate: describe a motion ("a bouncy walk cycle", "swing the sword"). It is added to the version you're viewing, and the first animation adds a skeleton for free. Every animation plays as a live preview in the list on the right of the canvas; click one to view it.
• Changing an animation: pick it in the animation list on the right, then open Revise. The tool switches to animation editing (the banner says which animation), and only the motion changes.

One prompt, one seed, four RD Plus styles:

Style families you'll see across models

Specialty styles go far beyond scenes — UI kits, weapons, sheets, levels, icons, and portraits:

Some of these settings in action:

→

A typical asset pipeline: generate → Inpaint the details you want changed → Background Remover → Color Reducer or Palette Converter to match your game's palette → download at original size.

Fix enlarged or softened pixel art

The standalone Pixel Art Fixer detects the underlying grid in enlarged, softened, AI-rendered, or compression-damaged pixel art and reconstructs the native pixels. Processing runs locally in your browser with the Rust/WASM engine, so the image is not uploaded for processing. Download the repaired art at its native 1× size or a crisp 4× preview.

Creating a style

1. Go to My Styles and click Create New Style.
2. Choose a template (RD Pro is the standard reference-image style) and give the style a name.
3. Upload 1–4 reference images that clearly show the style you want — consistent examples work far better than a mixed bag.
4. Optionally add a Prompt Modifier — extra text merged into every prompt. It must contain {prompt} where your per-generation prompt should be inserted.
5. Optionally force a color palette so every generation sticks to your colors.
6. Use Test Style inside the editor to run trial generations (charged like normal generations, results are temporary previews), tweak, and save.

Using and sharing

• Your styles appear under the My Styles tab in the generator's model panel — select one and generate as usual, with up to 9 reference images per generation.
• Choose who can use each style: Private (only you), Link (anyone with the link, kept out of Explore; make a new link any time to turn the old one off) or Public (listed in Explore with a cover image).
• Styles you add from Explore or a link stay connected to the original and pick up its creator's updates. Change anything and save, and it becomes your own remix, credited back to the original.

Images

• Browse, search and filter by model, style, size and shape; sort by trending, newest, likes, saves or copies.
• Like what you enjoy, save it for later, or Use settings to open the generator with that image's prompt, style, model and size.
• Share your own images with Share to Explore on the canvas. Shared images show your name and link to your collection; you can remove them any time. Images made with free usage are public.

Styles and remixes

• Every community style has a cover and the images people made with it. Add style puts it in your My Styles.
• Customize an added style and it becomes your remix, always credited to the original. Popular styles show their remix tree: every public remix, branching from the original.

1. Get an API key

• Go to Developer Tools and click the + button under API Keys. You can have up to 5 keys.
• The full key (starting with rdpk-) is shown once at creation — store it somewhere safe. The page only ever shows the prefix afterwards.
• API usage spends the same account balance as the website. Revoke a key any time from the same page.

2. Generate an image

Send a POST request to the inferences endpoint with your key in the X-RD-Token header:

import base64
import time
import requests

url = "https://api.retrodiffusion.ai/v2/inferences"
headers = {"X-RD-Token": "YOUR_API_KEY"}
payload = {
    "prompt": "A wizard's tower on a floating island, waterfalls",
    "prompt_style": "rd_fast__default",
    "width": 128,
    "height": 128,
    "num_images": 1,
}

start = requests.post(url, headers=headers, json=payload)
start.raise_for_status()
task_id = start.json()["task_id"]

while True:
    response = requests.get(f"{url}/tasks/{task_id}", headers=headers)
    response.raise_for_status()
    task = response.json()
    if task["status"] in ("pending", "running"):
        time.sleep(2)
        continue
    if task["status"] == "failed":
        error = task["error"]
        raise RuntimeError(f"{error['code']} (request {error['request_id']})")
    data = task["result"]
    break

for i, image in enumerate(data["base64_images"]):
    with open(f"output_{i}.png", "wb") as f:
        f.write(base64.b64decode(image))

print(f"Remaining balance: ${data['remaining_balance']}")

What else the API can do

• Every model, style, animation, and tileset workflow on the site — plus editing tools (from $0.001 per run) and the Pixel Fixer ($0.002 per call) — through a handful of endpoints.
• Free cost checks before generating, async job queueing, style discovery, and programmatic custom style creation.
• Low-Poly 3D models: generate, revise, animate, and export them from code.
• Assistant sessions: the Generation Assistant as an API. Your program chats, gets priced plans, and approves them.
• An MCP server so AI assistants and agents (Claude Code, Cursor, Codex, and others) can generate or repair pixel art directly — setup guides for every client live in the retro-diffusion-mcp repo on GitHub.

The full reference — every endpoint, every field, working examples, and a copy-paste context block for AI agents — lives on its own page:

• Every price is shown before you commit. The Generate button and every edit tool's Apply button display the exact cost of that action.
• What affects price: the model and the image count. Images are a flat price at any size: RD Fast $0.03, RD Plus and RD Mini $0.06 (Mini's fast styles $0.03), RD Pro $0.18. Animations and tilesets have flat per-run prices. Low-Poly 3D models are priced by size (see Low-Poly 3D).
• The Generation Assistant is free to chat with on the site; you pay only for the plans you approve.
• Top up by clicking your balance or visiting the Top Up page — packages from $5, with a bonus on larger amounts, paid through Stripe.
• Auto-refill can be configured under Payment Methods so long sessions and API workloads never stall.
• Failed generations are refunded automatically.
• Some free editing tools just require a small positive balance to use.
• Monthly invoicing — teams and enterprise customers can skip prepaid top-ups entirely and receive an invoice each month based on actual spend. Reach out on Discord or email support@retrodiffusion.com to get that set up.