The problem
Most AI image tools want you to learn their dialect before they will give you anything useful. Midjourney expects slash commands and weighted prompt syntax in Discord. DALL-E gives you one shot per request, and if the result is almost right you start over and hope. Firefly bundles the workflow inside a Creative Cloud subscription you may already be paying for under another line item.
Two failures keep showing up across all of them. The first is text rendering. Logos turn into glyph soup, signage in product mockups comes back as nonsense, and CJK characters fall apart. The second is iteration. You can describe the image perfectly on attempt seven, except the model has no memory of attempts one through six, so refinement is really starting over with extra context glued onto your prompt.
Pixparkle is what came out of trying to fix both problems in one product.
What Pixparkle is
Pixparkle is a chat-based AI image generator. You describe the image in plain language, see the result, and refine it through follow-up messages in the same conversation. The model keeps the context, so "make the lighting cooler" or "swap the logo color to lime green" works the way you would expect when talking to a designer.
Under the hood, Pixparkle runs on Google's Nano Banana model family plus a Flux Fast tier for quick drafts. You can switch between models mid-conversation without losing the thread:
- Flux Fast: 1 credit per image. Best for rapid volume work where speed beats polish.
- Nano Banana: 2 to 5 second draft generation at 1K resolution. 5 credits per image.
- Nano Banana 2: 4 to 6 second generation up to 4K, ultra-wide aspect ratios (8:1, 1:8, 4:1, 1:4), and an Image Search tool. 7 to 14 credits.
- Nano Banana Pro: Built on Gemini 3 Pro with a thinking stage that analyzes composition, lighting, and palette before generating. Best aesthetic quality and best text rendering. 10 to 20 credits.
A single chat session can drop from Pro down to Nano Banana for cheap variants, then climb back up to Pro for the final hero shot. You never have to re-explain the brief.
Capabilities that change the workflow
Accurate text rendering. Nano Banana Pro handles English, Chinese, Japanese, Korean, and several other languages. Headlines, signage, product label copy, and short paragraphs all come out readable. For e-commerce product cards, YouTube thumbnails, and educational diagrams, this removes the round trip through Photoshop where you would normally redo the text by hand.
4K output (3840x2160). Most consumer AI image tools cap at 1024 or 2048 pixels. Pixparkle generates up to 4K, which is the resolution you actually need for printed packaging, large-format posters, and crisp hero banners on retina displays.
Up to 14 aspect ratios. Square, portrait, landscape, widescreen, 9:16 reels, 16:9 thumbnails, and ultra-wide 8:1 or 1:8 banners. The available set depends on which model you pick. You stop cropping a single output across five platforms and generate the right ratio per surface from the same conversation.
Reference image editing. Drop in an existing photo and edit it through chat. Replace backgrounds, recolor elements, swap props, or apply a style direction. Useful when you already have a hero asset and need ten variants without rebuilding it from scratch.
Character consistency. Nano Banana Pro keeps the same character identity across multiple scenes. Generate a brand mascot once and place that mascot in a dozen different compositions without it morphing into someone else.
A typical chat looks like this
- You: "Product photo of a dark chocolate bar on a walnut surface, moody lighting, with the brand text TASTE THE SPARKLE on the wrapper in chartreuse foil."
- Pixparkle: returns the image in 4 to 6 seconds.
- You: "Cooler highlight from the upper right, and scatter a few yuzu segments around it."
- Pixparkle: returns a refined version.
- You: "Switch to Pro at 4K and add a SHOP THE DROP CTA at the bottom left."
- Pixparkle: returns the final version with sharp text and the requested layout.
There is no prompt template to memorize. If you can write the brief, you can drive the tool.
Who it lands with
E-commerce sellers running 50 to 500 SKUs. The credit math typically lands at a small fraction of studio photography for a comparable batch, and the white-background and lifestyle outputs are Amazon and Shopify spec compliant.
YouTube and short-form creators who used to spend two hours per thumbnail in Photoshop. A typical session covers three to four A/B variants in five minutes with the headline rendered directly inside the image.
Social media managers running multiple brand accounts. Generate the hero asset once, then ask for 9:16, 1:1, and 16:9 variants with consistent typography and color in the same chat.
Educators and technical writers producing illustrated explainers with accurate multilingual labels.
Pricing
Credit-based. You pay for what you generate.
| Plan | Price | Credits |
|---|
| Free | $0 | 10 |
| Pro Monthly | $9.90 / mo | 500 |
| Premium Monthly | $29.90 / mo | 2,000 |
| Pro Yearly | $99.90 / yr | 6,000 |
| Premium Yearly | $299.70 / yr | 18,000 |
No credit card to start. The 10-credit free tier works across all four models, so you can sanity-check the output quality on the models you actually care about before paying anything.
Stack notes for the engineers in the room
Pixparkle runs on Cloudflare Workers via OpenNext, Next.js 15 App Router on the frontend, Tailwind v4, and Drizzle ORM over Cloudflare D1. Image generation calls go out to Gemini and Replicate, with results stored in R2. Auth is NextAuth v5 with Google OAuth.
State for long-running generation jobs and the credit reservation system lives in Durable Objects. Queues handle async post-processing. The whole thing ships globally on the Cloudflare edge, which is most of why the latency feels closer to a chat app than to a typical AI image pipeline.
A public API is on the roadmap. If you are building something that needs chat-based image generation as a primitive and want early access, reach out in the comments.
Try it
Free tier is open at https://pixparkle.com/. No credit card required. 10 credits land in your account when you sign in with Google.
Feedback welcome in the thread. The things I am most curious about: what your text-rendering tests look like in non-Latin scripts, and whether multi-model switching inside a single chat matches how you actually want to iterate.
