⚡ Batch API Image Generator v2.5.0 (Rust + Slint)

Desktop batch image generator built with Rust + Slint. Generates AI images at configurable intervals with a powerful prompt randomizer engine, advanced Image-to-Image conditioning with up to 5 reference images, and Burst Generation for continuous non-stop generation. Supports 51 models across 5 API providers (xAI, Google, OpenAI, WaveSpeed, Kie.AI).

---

Supported Providers & Models

xAI (direct API)

Model	ID
Grok Imagine Image	grok-imagine-image
Grok Imagine Image Quality	grok-imagine-image-quality

Google (direct API)

Model	ID
Gemini 2.5 Flash Image	gemini-2.5-flash-image
Gemini 3 Pro Image	gemini-3-pro-image-preview

OpenAI (direct API)

Model	ID
GPT Image 1.5	gpt-image-1.5
GPT Image 1	gpt-image-1
GPT Image 1 Mini	gpt-image-1-mini
DALL-E 3 (legacy)	dall-e-3

WaveSpeed.ai (unified API — 31 models)

Text-to-Image

Model	Family
wavespeed-ai/flux-2-max/text-to-image	Flux 2
wavespeed-ai/flux-2-dev/text-to-image	Flux 2
wavespeed-ai/flux-2-flash/text-to-image	Flux 2
wavespeed-ai/flux-2-flex/text-to-image	Flux 2
wavespeed-ai/flux-kontext-max/text-to-image	Flux Kontext
wavespeed-ai/flux-kontext-pro/text-to-image	Flux Kontext
bytedance/seedream-v5.0-lite/text-to-image	ByteDance
bytedance/seedream-v4.5/text-to-image	ByteDance
bytedance/dreamina-v3.1/text-to-image	ByteDance
google/nano-banana-2/text-to-image	Google
google/nano-banana-pro/text-to-image	Google
alibaba/wan-2.6/text-to-image	Alibaba (T2I only)
alibaba/wan-2.7/text-to-image	Alibaba
openai/gpt-image-2/text-to-image	OpenAI via WaveSpeed
wavespeed-ai/qwen-image-2.0-pro/text-to-image	Alibaba
kwaivgi/kling-image-o3/text-to-image	Kuaishou
x-ai/grok-2-image	xAI
x-ai/grok-imagine-image/text-to-image	xAI
wavespeed-ai/flux-kontext-max/multi	Flux Kontext Multi
wavespeed-ai/flux-kontext-pro/multi	Flux Kontext Multi
wavespeed-ai/uno/text-to-image	UNO
wavespeed-ai/wan-2.2/image-to-image	WAN (I2I only, image singular)

Image-to-Image / Edit (I2I native endpoints)

Model	Field	Max Refs	Notes
x-ai/grok-imagine-image/edit	image singular	1	xAI images/edits format
google/nano-banana-2/edit	images array	5
google/nano-banana-2/edit-fast	images array	5	Faster variant
google/nano-banana-pro/edit	images array	5
bytedance/seedream-v5.0-lite/edit	images array	5
bytedance/seedream-v4.5/edit	images array	5
wavespeed-ai/qwen-image/edit	image singular	1	Verified against provider docs
wavespeed-ai/flux-2-klein-4b/edit	images array	5
alibaba/wan-2.7/image-edit	images array	3	Provider caps at 3
openai/gpt-image-2/edit	images array	5	resolution 1k/2k/4k

Kie.AI (unified API — 11 models) (New in v2.5)

Upload → createTask → poll recordInfo. Every model's aspect-ratio and size fields are verified against Kie.AI's own docs: sending "auto" to a model that does not accept it returns HTTP 500, and omitting a required field returns This field is required. Both were hit in real use and both are now pinned.

Model	T2I	I2I	Reference field	Max refs
GPT Image 2	✅	✅	input_urls	5
Nano Banana 2	✅	✅	image_input	5
Nano Banana Pro	✅	✅	image_input	5
Nano Banana 2 Lite	✅	✅	image_urls	5
Seedream 4.0	✅	—	—	—
Seedream 4.0 Edit	—	✅	image_urls	5
Seedream 5.0 Pro	✅	✅	image_urls	5
Seedream 5.0 Lite	✅	—	—	—
Grok Imagine	✅	✅	image_urls	2
Flux.2 Pro	✅	✅	input_urls	5
Qwen Image 3.0	✅	✅	image_urls	3

---

Features

🎨 Three Switchable Skins (New in v2.5)

Dark · Light · Cyberpunk, switchable from the header and identical on Windows, macOS and Linux.

std-widgets.slint takes its palette from the desktop, which is why the app was unreadable under a light Ubuntu theme. Slint's Palette global only lets you write color-scheme at runtime — its colours are read-only — so three skins required a custom widget set. ui/widgets.slint provides it; every colour derives from a single Theme.skin integer.

🎰 Super Randomizer (New in v2.5)

Instead of picking the boxes yourself, each generation draws a random number of categories (1 to all 21) and a random subset of them. It does not replace the manual mode — it sits next to it.

Your manual selection is snapshotted while Super is on, so turning it off gives you back the combination you had, and closing the app never saves a random draw as if it were your choice.

📝 Prompt Bank (New in v2.5)

Save up to 5 prompts and optionally have one picked at random per generation. The Randomizer and Super Randomizer then apply on top of whichever was drawn.

🔔 Desktop Notifications (New in v2.5)

XDG/D-Bus on Linux (Wayland and XWayland alike — the transport is D-Bus, not the display server), WinRT toasts on Windows, UNUserNotificationCenter on macOS.

Per-event switches for success, server timeout and content-policy rejection. Success is off by default: in a long Burst it would be hundreds of alerts.

Diagnostics without opening the window:

./xai-imagine-generator --test-notification

Prints the real result of the call plus the environment (session D-Bus, which daemon answers, desktop, session type) — because "I see no notifications" has two very different causes and the window cannot tell them apart.

💾 Persisted Preferences (New in v2.5)

Skin, output folder, model, interval, mode, every randomizer switch, the prompt bank and the notification settings survive a restart. Stored in the platform's config directory (~/.config/batch-image-generator/config.json on Linux).

⚡ Burst Generation (New in v2.4)

Fires generation requests continuously with zero interval — as soon as one image finishes, the next starts immediately. Perfect for rapid iteration. Uses the same model, prompt and reference images as the configured loop.

🖼 Image-to-Image with up to 5 Reference Images (Expanded in v2.4)

• Upload up to 5 reference images (PNG/JPG/WEBP) to guide generation or directly edit existing pictures
• Choose between Style Reference (loose guidance) and Direct Edit (strict adherence to content)
• Each slot has a distinct color indicator (🔵 blue, 🟢 green, 🟣 purple, 🟠 orange, 🟡 yellow)
• Dynamic routing sends the correct field (image singular vs images array) per model's API contract
• Encodes images directly as Base64 data URIs — no separate upload step

🖥 Output Resolution Selector (New in v2.4)

Choose output resolution for all WaveSpeed models directly from the UI: Auto (model default), 1k (1024), 2k (2048), or 4k (4096). For GPT Image 2 Edit the resolution field is sent natively; for all other WaveSpeed models the size field is mapped accordingly.

🔄 Batch Loop

Configurable interval (10–600 seconds) between generations. Status countdown shows remaining time. Three control buttons:

• ▶ Iniciar Loop (Start Loop) — start the batch loop at the configured interval
• ⚡⚡ Burst — continuous generation with no pause
• ■ Detener (Stop) — stop either mode

The UI ships with Spanish labels; this README glosses them in English where they appear.

🎲 Prompt Randomizer

• Mode A — Write your own prompt; the randomizer injects additional details (outfit, lighting, pose, camera, nails, expression, etc.)
• Mode B — Fully auto-generated prompts from curated pools and theme presets
• 20+ injectable categories: nails, orientation, expression, outfit, legwear, environment, atmosphere, pose, lighting, camera, accessories, makeup, body type, age vibe, color palette, time of day, weather, background props, material, motion, rare details
• Smart prompt reinforcement — preserves the user's base prompt subject so AI models don't ignore it
• Advanced Regex Engine — strict word-boundary algorithms target precise descriptors without overwriting the original prompt's core content

🔑 Three API Key Slots

Separate keys for xAI/Google/OpenAI (direct), WaveSpeed.ai and Kie.AI.

Keys are not written to config.json. Storing them in plain text would be a security regression; the right place is the system keyring, which is pending. Everything else is persisted.

🖥 Cross-Platform

Windows, macOS, Linux.

---

Build

cargo build --release

The binary will be at target/release/xai-imagine-generator (or .exe on Windows).

Requirements

• Rust 1.70+
• An API key from at least one provider:
  • xAI Console for Grok models (direct)
  • Google AI Studio for Gemini / Nano Banana (direct)
  • OpenAI Platform for GPT Image / DALL-E (direct)
  • WaveSpeed.ai for Flux, Seedream, WAN, Grok via WaveSpeed, and 30+ more models
  • Kie.AI for GPT Image 2, Nano Banana, Seedream, Flux.2 and Qwen

Prebuilt binaries for Windows, macOS (Intel and Apple Silicon) and Linux are attached to every release. On Linux and macOS, chmod +x the binary before running it.

---

Architecture

src/
  main.rs        — UI wiring, callbacks, countdown timer, generation loop, burst mode
  models.rs      — single source of truth: 51 models, provider, endpoints, image field
  api.rs         — multi-provider HTTP client (OpenAI-compat · WaveSpeed · Kie.AI)
  randomizer.rs  — prompt engine (Mode A: inject · Mode B: generate · Super Randomizer)
  pools.rs       — randomization pools (styles, outfits, environments, …)
  config.rs      — preferences persisted across runs (skin, folder, model, switches)
  notify.rs      — cross-platform desktop notifications
  util.rs        — char-safe string helpers (byte slicing used to panic on UTF-8)
ui/
  theme.slint    — three skins; every colour derives from one `Theme.skin` int
  widgets.slint  — custom widget set, so the look does not follow the OS palette
  main.slint     — window layout

---

Technical Notes

Model Table — single source of truth (rewritten in v2.5)

Until v2.4 the endpoint and the name of the reference-image field were guessed from the model identifier with a cascade of contains() / ends_with() checks, and the same information was duplicated by hand in three places (MODEL_CATALOG, the Slint dropdown, and the routing code). Nothing kept them in sync, and it broke: the // ← was missing comment next to ends_with("/edit-fast") is the scar.

Kie.AI made it untenable — its models use three different names for the same reference-image field (input_urls, image_urls, image_input) with no rule derivable from the identifier.

Every model is now declared once in src/models.rs:

ModelSpec {
    label: "WaveSpeed — WAN 2.2",
    provider: ImageProvider::WaveSpeed,
    t2i_id: None,
    i2i_id: Some("wavespeed-ai/wan-2.2/image-to-image"),
    ref_field: RefField::WsImage,   // verified against provider docs
    max_refs: 1,
    aspect: "",
    size_style: SizeStyle::WsSize,
}

The dropdown labels, the routing and the UI hints all come from that table, so the window cannot show one model and send another. Tests pin the exact request body per model, so a wrong field is caught in half a millisecond instead of by an HTTP 400 that costs credits.

WaveSpeed Polling Flow

POST to submit → sync mode waits for completion → download image from CDN URL. Falls back to polling if sync mode times out (180s).

Resolution Defaults

• ByteDance (Seedream, Dreamina): 1920×1920 (minimum required by provider)
• All other WaveSpeed models: 1024×1024

xAI / Google / OpenAI

Use OpenAI-compatible endpoints returning b64_json. Output saved directly from the response.

---

Changelog

v2.5.0

• New: three switchable skins (Dark / Light / Cyberpunk) on a custom widget set, so the look no longer follows the desktop palette.
• New: Kie.AI as a fifth provider — 11 models, upload → createTask → poll.
• New: Super Randomizer — a random number of categories drawn per generation.
• New: prompt bank of 5 slots with optional random pick per generation.
• New: cross-platform desktop notifications, with --test-notificacion for diagnosing them from a terminal.
• New: preferences persisted between runs.
• Refactor: src/models.rs is now the single source of truth for all 51 models; the contains() cascade that guessed endpoints and field names is gone, and with it the class of bug it kept producing.
• Fix: four UTF-8 panics — &s[..n] on a multibyte boundary crashed the app on any accented prompt.
• Fix: Burst overwrote images generated within the same second.
• Fix: Stop did not cancel the in-flight request — the app said STOPPED and a billed image appeared minutes later.
• Fix: WAN 2.2 and Qwen Image Edit sent images (array) where the provider requires image (single) — HTTP 400 on every request.
• Fix: a CDN hiccup threw away an already-paid generation; the download now retries transient failures and keeps the URL if it gives up.
• Fix: the log grew O(n²) and was never trimmed.
• Fix: one HTTP client is shared instead of one per request; regexes compile once.
• CI: builds and tests on Linux, Windows and macOS on every push; the release workflow no longer uploads two assets with the same name.

v2.4.0

• New: GPT Image 2 Text-to-Image (openai/gpt-image-2/text-to-image) via WaveSpeed added to model catalog.
• New: GPT Image 2 Edit (openai/gpt-image-2/edit) via WaveSpeed — I2I with images array + native resolution field.
• New: Output resolution selector in UI — Auto / 1k / 2k / 4k. Applies to all WaveSpeed models.
• New: resolution_to_size helper maps UI selection to size string (1024*1024 / 2048*2048 / 4096*4096).
• New: output_resolution: &str parameter threaded through generate_image, generate_wavespeed, generate_wavespeed_i2i.
• Fix: GPT Image 2 T2I uses a schema without size field (only prompt + aspect_ratio); handled with a dedicated branch.

v2.3.0

• Fix: x-ai/grok-imagine-image/edit was sending images (array) — API requires image singular (xAI images/edits format). Added is_grok_edit flag.
• Fix: google/nano-banana-2/edit-fast (/edit-fast suffix) was not included in needs_images_array — fell through to singular field. Added ends_with("/edit-fast") to the condition.
• Fix: alibaba/wan-2.6 with I2I active tried to hit non-existent /image-to-image endpoint. Now returns a descriptive error directing to wan-2.7/image-edit.
• Fix: Models already carrying /edit or /edit-fast in their name were having the suffix appended again (…/edit/edit). Added both to the "already I2I" early-return check.
• New: Burst Generation mode — continuous generation with zero pause between requests.
• New: 9 native I2I/edit endpoints added (Grok Edit, Nano Banana 2 Edit/Edit-Fast/Pro, Seedream v5 Lite/v4.5 Edit, Qwen Edit, Flux 2 Klein Edit, WAN 2.7 Image-Edit).
• New: UNO, WAN 2.2 I2I, Flux Kontext Multi, Kling O3, Qwen Image 2.0 Pro, Dreamina v3.1, WAN 2.7 added to MODEL_CATALOG (40 total).
• New: Reference image slots expanded from 2 to 5 (slots 3–5 with purple/orange/yellow indicators).
• New: generate_image signature changed to ref_images: &[(String, String)] for variable-count multi-image support.
• New: WaveSpeed body construction switched from typed structs to serde_json::json! macro for dynamic field routing without lifetime complexity.
• UI: Two-row control button layout — Loop + Burst + Stop on top row; Generate 1 on bottom row.
• UI: Dynamic I2I info note adapts message based on number of loaded reference images.

v2.2.0

• Universal Image-to-Image routing across all WaveSpeed endpoints.
• Stable asynchronous countdown batch loop (no duplicate simultaneous requests).
• Advanced Regex Engine for the Randomizer (word-boundary checks for nails, expressions, etc.).
• Expanded do_nails vocabulary (3D, glossy, claw-like with structural shape + color).

v2.1.0

• Image-to-Image conditioning (2 reference image slots).
• Style Reference vs Direct Edit modes.
• Base64 encoding pipeline.

v2.0.0

• Initial Rust + Slint rewrite.
• Multi-provider support (xAI, Google, OpenAI, WaveSpeed).
• Prompt Randomizer (Mode A & B).
• Configurable batch loop.

---

License

Released under the MIT License — see LICENSE for the full text.

Copyright (c) 2026 Eric .V

You are free to use, copy, modify, merge, publish, distribute, sublicense and sell copies of this software, provided the copyright notice and the permission notice are included in all copies or substantial portions of it. The software is provided "as is", without warranty of any kind.

Third-party services

The MIT licence covers this application only. Images you generate are subject to the terms of whichever provider produced them — xAI, Google, OpenAI, WaveSpeed.ai or Kie.AI — including their content policies and their rules on commercial use. Check each provider's terms before publishing or selling generated output.

---

Author

Eric Valls Gramunt — @AcidClawX41

The copyright line above is quoted verbatim from LICENSE. If you want your full name there, edit LICENSE and this section will need to match.