Generate Images on Perchance AI Text-To-Image

Purpose

Drive Perchance's free, no-signup AI Text-To-Image generator to produce images from a text prompt, apply the available style and rendering options, wait for the batch to finish, and extract each result's metadata (expanded prompt, guidance scale, per-image seed) and inline image bytes. Optionally save the best result(s) to the generator's private gallery. Useful for bulk/iterative image generation where you want full control over style, shape, batch size, and per-image seeds. This is a generation task — nothing is purchased and there is no account; it is free and unlimited.

When to Use

Generate one or more AI images from a natural-language prompt, free and without login.
Bulk-generate a batch (4–32 images) and pick/save the best candidate(s).
Exercise specific creative controls: art style preset, aspect ratio (shape), image count, and prompt modifiers (camera shot, color, effect, genre).
Capture the exact generation parameters (expanded prompt, guidance scale, seed) for reproducibility or downstream analysis.

Workflow

The recommended (and only viable) method is browser: the generator is powered by Perchance's in-browser text-to-image-plugin running inside nested iframes — there is no documented public REST API or query-string deep-link that triggers generation, so scripted browsing of the embedded UI is the correct path.

Start a stealth remote session. The perchance.org domain sits behind Cloudflare (the bare homepage returns HTTP 403). Create the session with --verified --proxies (i.e. browse open … --remote against a verified + residential-proxy session). Navigate to https://perchance.org/ai-text-to-image-generator. Wait ~3s, then confirm the page title is AI Image Generator (free, no sign-up, unlimited).
Snapshot to get cross-frame refs. Run browse snapshot. It flattens the nested iframes via CDP and exposes stable refs in the [1-*] namespace (the generator lives inside #outputIframeEl, NOT on the top page). Key controls:
- [1-3] — Description / prompt textbox (pre-filled with an example; just browse fill over it).
- [1-4] — 🎨 Art Style <select> (default "Painted Anime"; 80+ options: Cinematic, Casual Photo, Digital Painting, Concept Art, "No style", Studio Ghibli, Pixel Art, Watercolor, Manga, …).
- [1-5] — 🖼️ Shape <select>: Square (default) / Portrait / Landscape.
- [1-6] — 🔢 How many? <select>: 4 (default) / 6 / 8 / 16 / 32 / 2.
- [1-9] — "✨ generate" button.
- Optional prompt-modifier selects that appear near the prompt: "Add shot…", "Add color…", "Add effect…", "Add genre…".
Fill inputs. browse fill [1-3] "<prompt>". Optionally browse select [1-4] "<Art Style>", browse select [1-5] "<Shape>", browse select [1-6] "<count>" (pass the visible option label).
Generate. browse click [1-9].
Wait and poll for completion. Each requested image renders in its own nested "Perchance Image Generation Embed" iframe. Snapshots first show placeholders ("Waiting…" / "Preparing…"); an image is finished only when its image node's accessible name contains a seed= value. Poll with browse snapshot every ~20s. Budget 30–90s per batch — large batches (16/32) take longer. Do not stop after the first snapshot or you will capture only a partial batch.
Extract results. For each finished image node, parse its accessible name: prompt=<expanded prompt + style keywords> negativePrompt=<…> guidanceScale=<n> seed=<n>. The raw pixels are available as an inline data:image/jpeg;base64,… src on the embed's image element (there is no external CDN URL).
(Optional) Save the best. Click the 🛡️💾 button beneath the chosen image to save it to the private gallery (viewable via the 🛡️ show private gallery button). Pick based on prompt intent; do not blindly save all.
Screenshot the results grid as evidence.

Site-Specific Gotchas

Cloudflare 403 on the domain. A bare HTTP request to perchance.org returns 403. Use a stealth remote session (verified + residential proxies). In testing the generator subpage occasionally loaded on a plain Browserbase remote session too, but this is not reliable — keep stealth on to avoid intermittent challenges.
Nested iframes — three layers deep. The generator is NOT on the top page. It lives in #outputIframeEl (a *.perchance.org subdomain iframe), and each result renders in a further-nested "Perchance Image Generation Embed" iframe. Use browse snapshot (which flattens frames via CDP into [1-*], [60-*], [61-*]… refs) rather than trying to read iframe.contentDocument (cross-origin → null).
Image bytes are inline base64, not URLs. Finished images are data:image/jpeg;base64,… blobs painted into sandboxed embed canvases. There is no external/CDN image URL to grab. Capture the data URI from the src, or read the metadata from the image node's accessible name.
All generation metadata is in the accessibility name. Each finished image node encodes prompt=, negativePrompt=, guidanceScale= (defaults to 7), and seed= (random per image). This is the most reliable structured output.
Selecting an art style mutates the expanded prompt. Choosing e.g. "Cinematic" auto-appends keywords (75mm, Technicolor, Panavision, cinemascope, HDR, "world-class cinematic masterpiece", …) to the prompt visible in the image node name. Expected, not a bug.
Default batch is 4. Leave "How many?" modest; 16/32 dramatically increase wait time. Partial-batch capture is the most common failure — poll until all requested images carry a seed=.
Ad iframes pollute the snapshot. ads.perchance.org, doubleclick, pubmatic, etc. add dozens of irrelevant refs. Ignore everything outside the generator frames.
No login / no payment / unlimited & free. No auth wall to clear; never attempt to log in.

Expected Output

{
  "success": true,
  "prompt": "a cozy cabin in a snowy pine forest at sunset, warm glowing windows",
  "style": "Cinematic",
  "settings": {
    "shape": "Square",
    "num_images": 4,
    "guidance_scale": 7,
    "negative_prompt": ""
  },
  "images": [
    {
      "seed": 121257895,
      "expanded_prompt": "a cozy cabin in a snowy pine forest at sunset, warm glowing windows, cinematic shot, dynamic lighting, 75mm, Technicolor, Panavision, cinemascope, sharp focus, fine details, 8k, HDR, ... cinematic color grading, depth of field.",
      "src": "data:image/jpeg;base64,/9j/4AAQSkZ...",
      "saved": false
    },
    {
      "seed": 272163316,
      "expanded_prompt": "a cozy cabin in a snowy pine forest at sunset, warm glowing windows, cinematic shot, ...",
      "src": "data:image/jpeg;base64,/9j/4AAQSkZ...",
      "saved": true
    }
  ],
  "image_count": 2,
  "error_reasoning": null
}

Failure / partial shapes:

// Still rendering at capture time — poll longer
{ "success": true, "image_count": 2, "images": [ { "seed": null, "src": null, "alt": "Waiting... (not yet complete)" } ], "error_reasoning": null }

// Blocked by Cloudflare (no stealth)
{ "success": false, "image_count": 0, "images": [], "error_reasoning": "Cloudflare 403 / challenge page — recreate session with --verified --proxies" }