🔥 Emberforge

Why "Emberforge"? It started as a plain image generator and grew into a forge — a desktop workbench where you don't just press "generate", you heat raw ideas into shape. It melts together many engines (cloud APIs and your own local ComfyUI/Ollama rig), hammers prompts into structured, spatially-aware Ideogram V4 captions, and lets variants evolve through a mutation/breeding engine. Forge for the craft and heat; ember nods to the home GPU box ("fireEngine") that does the local rendering. (The repo and project keep their original ImageGenerator.MAUI identity; only the app name, the shipped binary — Emberforge.exe — and the icon are Emberforge.)

A Windows desktop (.NET MAUI) image-generation workbench that routes one prompt to many backends — Replicate (Flux, Replicate-hosted OpenAI, Google, Ideogram V4), Pollinations.ai (Flux, Zimage, Qwen, plus anything their /models lists), and ComfyUI (your own local/LAN server, any API-format workflow becomes a model) — and adds an LLM-assisted prompt pipeline on top.

What makes it interesting

• Three providers, one app. Pick a model and the request routes itself; each provider's token lives in its own OS-secure slot. New Replicate/Pollinations models appear via Refresh Models without recompiling.
• Your own ComfyUI as a first-class backend. Drop an API-format workflow JSON in a folder and it becomes a selectable model. The app patches prompt / seed / resolution / checkpoint / quality-preset into the graph and streams live per-step progress over the server WebSocket. No cloud, no cost beyond your GPU. → details
• Describe an idea → prompt (Claude or Ollama). Type a plain idea; Claude Opus/Sonnet or a local Ollama model writes a polished prose prompt that works with any model, and optionally maps it to a schema-valid Ideogram V4 structured caption. → details
• Visual structured-prompt editor. Build Ideogram V4's json_prompt in a three-pane workspace with an inspector, responsive canvas, and live JSON output — drag/resize element boxes on the 0–1000 grid, set style/palette, and Enrich from layout: a geometry-grounded LLM rewrites each element's description to reflect its real spatial relationships (what it rests on, sits beside, or is behind). → details
• Mutation & breeding engine. Turn any structured caption into a batch of one-change variants along a LOOK (style) or SCENE (composition) axis — fully deterministic, seeded, offline, no key. Or switch on AI mode to steer mutations and breed from your favourites via Claude (Sonnet/Opus) or a local Ollama model. → details
• Post to CivitAI. One checkbox publishes a finished image — optionally into a specific model's gallery with structured generation data attached. → details

Everything else, briefly: a self-contained single-file exe (no installer); a concurrent generation queue with per-job cancel; batch a .txt of prompts sequentially; an in-app gallery (sortable, live-watched, multi-select, copyable metadata, use as input, remix); every saved file embeds its full prompt + parameters (PNG Comment / EXIF UserComment) so any output is reproducible; per-provider tokens in Windows SecureStorage (DPAPI); persisted prompt/model/resolution/window bounds; and a diagnostics log at %LOCALAPPDATA%\Emberforge\app.log (NLog, rolls at 5 MB).

📸 Screenshots

🖥️ Getting started

Download. Grab Emberforge.exe from the latest release — self-contained, no installer or runtime prerequisite. It ships unsigned, so on first run Windows SmartScreen shows "Windows protected your PC" → More info → Run anyway. (Or build from source.)

Tokens & providers. The Settings page has one API-Tokens picker with an independent, DPAPI-secured slot per provider — set whichever you use and switch models freely on the main page:

Provider	What it's for	Where to get it
Replicate	Flux · OpenAI gpt-image · Google nano-banana-2 · Ideogram V4	replicate.com/account/api-tokens
Pollinations	flux / zimage / qwen-image etc. — works anonymously (rate-limited); a token raises the limit	auth.pollinations.ai
Anthropic	Claude tiers for Describe an idea + AI mutation/enrichment (never image gen)	console.anthropic.com
CivitAI	publishing finished images (Media-Write scope)	civitai.com/user/account
ComfyUI	a server base URL (and optional Authorization header) instead of a token	your own server

Reproducibility note: Pollinations honours seed only on flux, zimage, seedream, klein, seedance, nova-reel; other models ignore it.

Your own ComfyUI server

The ComfyUI provider talks to a ComfyUI instance you run (localhost or LAN — start it with --listen for LAN, and don't expose a bare port to the internet). The app never builds graphs; you bring your own. In ComfyUI use Workflow → Export (API) and save the JSON into your data folder's comfy-workflows\ subfolder (default Pictures\Emberforge\comfy-workflows\, or <output folder>\comfy-workflows\ when you've set one in Settings) — every file there becomes a model entry, re-scanned on launch / Refresh Models. (Normal Ctrl+S saves are the UI format and can't be queued; the app tells you to re-export.)

At submit time the app patches the graph: your prompt into the lowest-id CLIPTextEncode (or, in structured mode, into Ideogram4PromptBuilderKJ / a JSON-literal CLIPTextEncode); re-rolls every literal seed / noise_seed; writes aspect-ratio + megapixels into a ResolutionSelector if present; offers a checkpoint picker for a baked CheckpointLoaderSimple or single literal UNETLoader (multi-UNET pairings keep their models); offers a quality-preset picker for a single CustomCombo; and expands %date:…% filename tokens app-side. Minimum a workflow needs: a CLIPTextEncode with literal text and a SaveImage.

comfy-workflows/Ideogram4-Sample.json is a ready-to-use Ideogram 4 workflow using stock ComfyUI nodes only; it needs the usual Ideogram 4 model files on the server (ideogram4_fp8_scaled + ideogram4_unconditional_fp8_scaled, qwen3vl_8b_fp8_scaled text encoder, flux2-vae). Canceling in the app drains pending jobs and interrupts the active render when it's the app's own.

Upscaling. A workflow containing a LoadImage node is an image-to-image workflow: the app shows the Input Image card for it, uploads your image to the server (/upload/image) and patches the stored name into every LoadImage. When such a workflow's file name contains upscale, it also becomes the target of the Upscale after render checkbox (ComfyUI models only): the finished render is fed straight into it and both files are kept, the second one with an _upscaled suffix. comfy-workflows/Upscale-Sample.json is a ready-to-use tiled Ultimate SD Upscale 2× pass (SDXL zavychromaxl + TTPlanet tile ControlNet + 4x_foolhardy_Remacri); to upscale an existing image, select it as the model and use Use as input on any gallery image.

Krea-2 LoRAs. A workflow whose CLIPLoader has type: "krea2" gets a LoRA picker in the Output card. The picker reads the configured host's native GET /models/loras catalog (including subfolders), offers None, and remembers the selection per workflow. At queue time Emberforge injects ComfyUI's built-in LoraLoaderModelOnly with the selected model strength; the workflow JSON on disk is never modified. Only choose LoRAs trained for Krea-2—the ComfyUI catalog contains every file in models/loras and does not label model compatibility.

Describe an idea (Claude or Ollama)

Describe an idea… runs two passes: Pass 1 always turns a plain-English idea into a polished prose prompt (good for any model); Pass 2 optionally maps that onto a schema-valid Ideogram V4 JSON caption. The result card lets you copy the prose, use it as the prompt, or use the JSON. Pick Claude Opus, Claude Sonnet, or Local (your configured Ollama server/model). Anthropic tiers are billed per pass; Local needs no token and quality depends on the installed model. Power users can override the bundled clean-room instructions with private vpe-prompt.md / system-prompt.md files in your data folder's prompt-builder\ subfolder (default Pictures\Emberforge\prompt-builder\) (read fresh, never enter the repo — an open-core split).

Mutation & enrichment

From any Ideogram V4 structured caption (hand-built, from Describe an idea, or remixed from a saved image):

• Mutate it into a batch of one-change variants — LOOK (style/ornament; 24 built-in style fragments) or SCENE (background, placement, element presence; aspect-ratio-aware bbox moves) — one axis per run so every difference is traceable. Deterministic, seeded, offline by default. Per-element slot tags decide what a mutation may touch; the subject's identity is always preserved.
• AI mode (optional, needs an Anthropic key or a local Ollama endpoint) lets you steer mutations in plain language and breed new captions from the variants you liked.
• Enrich from layout (in the structure editor) rewrites each element's description to read its spatial place in the scene. A deterministic pass computes the geometry (relative position, support/contact, background band, overlap) and the LLM turns it into natural relational prose — deciding front/behind from the geometry and each description's own wording, never from list order. You preview before/after per element and accept or discard.

Render a batch at one fixed seed (so only the change differs), pick winners in the gallery, and promote one with Remix.

Posting to CivitAI

Check Post to CivitAI and a finished image is published the moment it's saved — optionally into a model's gallery (paste the model URL or version id) with prompt/seed/model attached as structured metadata (CivitAI ignores file metadata on API uploads, so this is how generation data reaches a post). It runs after the file is on disk and never blocks or fails the generation; the checkbox resets off each launch. Uses CivitAI's official MCP server for the upload plus one direct API call for the post.

Costs

Replicate is pay-as-you-go, roughly $0.003–$0.05 per image by model (pricing). Pollinations free models stay free; paid_only models are filtered out of the picker. Anthropic applies only when you pick a Claude tier for the prompt builder / AI mutation, billed per call.

🚀 Building from source

Needs the .NET 10 SDK + the MAUI workload (VS 2022 17.12+ or Rider), Windows 10 1809+.

git clone https://github.com/MR-444/ImageGenerator.MAUI.git

Open ImageGenerator.MAUI.sln, restore, build, run. For the self-contained single-file release exe, run pwsh ./publish.ps1 from the repo root. Tests: dotnet test (1368 tests — provider payloads, the ComfyUI patcher, catalog filtering/persistence, the V4 structured-prompt model + validator, the deterministic mutation operators and the RegionGraph geometry, the LLM seams via fakes, CivitAI posting, gallery + UI-state persistence).

🛠️ Stack

.NET MAUI 10 (Windows) · CommunityToolkit.Mvvm · Refit + System.Text.Json · Microsoft.Extensions.Http.Resilience (Polly) · SixLabors.ImageSharp 3 · raw HttpClient for the Anthropic/Ollama LLM calls · xUnit + Moq + FluentAssertions.

⚠️ Known issues

• Error-path UX (401s, mid-generation network drops, 5xx/quota) isn't exhaustively audited — happy path, Cancel, and concurrent jobs work as expected.
• Gallery has no in-app delete/rename/search yet (manage files in Explorer; the watcher picks changes up within ~1 s).
• Pollinations kontext (image-to-image) has no input-image plumbing — use a Replicate Flux model for image-prompted edits.

Hitting something? Open an issue with the status text from %LOCALAPPDATA%\Emberforge\app.log (i.e. C:\Users\<you>\AppData\Local\Emberforge\app.log) and what you were generating.

🔭 Roadmap

Directions under consideration (a small side project, not commitments): one-click upscale (via a Replicate/ComfyUI upscaler) · saved parameter presets · finer per-element control in the enrichment preview.

📄 License

MIT — see LICENSE. Initial work by Silmas. Thanks to the .NET MAUI team and the image-generation API providers.