Image Generation MCP Servers — DALL-E, Stable Diffusion, ComfyUI, Flux, Midjourney, and 45+ More

Image generation is one of the most natural use cases for MCP integration. Instead of switching between generation platforms, tweaking parameters in web UIs, or managing API keys across services, an AI agent can generate, edit, upscale, and manipulate images conversationally. The ecosystem has responded with 45+ MCP servers spanning every major generation platform.

April 2026 update: OpenAI released gpt-image-2 on April 21, 2026 — with native reasoning built into the generation pipeline (it plans before rendering), up to 4K resolution, and ~99% text-rendering accuracy in English (90%+ across CJK, Hindi, Bengali, and Arabic scripts). OpenAI describes it as roughly 2x faster than gpt-image-1. MCP servers are already adding support. ComfyUI’s lead server hit 398 stars with a major v1.0 architectural rewrite. Adobe Firefly finally has an MCP server (closing a key gap from our initial review), though adoption is early. (Correction, 2026-08-15 audit: earlier versions of this page described gpt-image-2 as running on a “GPT-5.4 backbone” and cited a “3-5x faster” figure inconsistently alongside a separate “2x faster” claim. Neither OpenAI’s own model docs nor its announcement thread name a specific backbone model or a 3-5x speed figure — those numbers weren’t supportable and have been corrected/cut.)

The landscape splits into three tiers: local generation (ComfyUI, Stable Diffusion WebUI, local Flux models running on your GPU, Draw Things on Mac), API-based generation (OpenAI/DALL-E/gpt-image-2, Stability AI, Google Imagen/Nano Banana, Midjourney via proxies), and platform aggregators (Replicate, fal.ai, Together AI, PiAPI — services that provide access to hundreds of models through a single API).

The headline findings: ComfyUI has the richest MCP integration with four servers and up to 40+ tools for workflow-based generation. Stability AI’s API wrapper is the most polished single-platform server with 11 chainable tools and new SD 3.5 support. fal.ai provides the broadest model access — raveenb/fal-mcp-server alone claims 600+ models available through a single MCP server. Google Gemini/Imagen is surging with shinpr/mcp-image at 152 stars and auto-prompt optimization powered by Nano Banana 2/Pro models. gpt-image-2 is the new quality benchmark — lansespirit/image-gen-mcp already supports it alongside gpt-image-1.5 and imagen-4-ultra. Midjourney opened an Enterprise API application process in mid-2025, but as of this audit it remains an application/waitlist process, not a public launch — developer access is still restricted. Adobe Firefly now has an MCP server via msabramo/python-firefly, though it’s early (2 stars).

For our review of EverArt’s reference MCP server (one of Anthropic’s original examples), see our dedicated EverArt review. For photography-focused image tools (editing, stock photos, EXIF), see our Photography MCP review.

ComfyUI (4 servers)

ComfyUI — the node-based Stable Diffusion interface — has the most sophisticated MCP integration in the image generation space. Its workflow-based architecture maps naturally to MCP tools: each workflow becomes an invocable function.

joenorton/comfyui-mcp-server (398 stars, up from 294) is the standout and recently shipped a v1.0 architectural rewrite. The new version replaces WebSocket connections with streamable HTTP transport, adds explicit job management (queue status, cancel, asset metadata), and introduces publishing tools for asset distribution. It still auto-discovers ComfyUI workflows and exposes them as MCP tools. The iterative refinement feature lets agents generate, evaluate, and regenerate with adjusted parameters. The v1.0 migration is a breaking change — users of earlier versions should consult the migration notes.

shawnrushefsky/comfyui-mcp (18 stars, up from 8) takes the broadest approach with 38+ tools across 9 categories (setup, templates, generation, composition, discovery, task management, agent memory, preferences, SVG/font utilities) and 70+ example workflows covering images, video, audio, and 3D generation. It’s self-configuring — it introspects the connected ComfyUI instance to determine available models and capabilities. Docker support makes deployment straightforward.

artokun/comfyui-mcp (589 stars, up sharply from 9) is interesting for its agent-centric design: it now includes 4 autonomous agents (explorer, debugger, optimizer, researcher) for image generation, video creation, and model management, plus 11 slash commands for common operations. It tracks all generations in SQLite for history and auditability.

The ComfyUI subcategory is the most feature-rich in the entire image generation MCP space. The trade-off is complexity — you need a running ComfyUI instance with models installed, which means significant disk space (tens of GB) and ideally a capable GPU.

Stability AI / Stable Diffusion (4 servers)

tadasant/mcp-server-stability-ai (84 stars, up from 81) is the best single-platform API wrapper in the entire image generation category. It recently added Stable Diffusion 3.5 and Stable Image Ultra support with metadata logging for tracking requests. Its 11 tools are chainable — generate an image, remove its background, upscale it, then apply relighting, all in a single conversation. The control tools (sketch, style, structure) let agents apply image-to-image transformations with precise control. It uses Stability AI’s cloud API, so no local GPU needed, but API costs apply.

Ichigo3766/image-gen-mcp and boxi-rgb/sd-webui-mcp both bridge MCP to a local AUTOMATIC1111/Forge WebUI instance. If you’re already running SD locally, these let agents generate images using your installed models and LoRAs without API costs. The trade-off is that you need the WebUI running as a server.

DALL-E / OpenAI (7+ servers)

Breaking: gpt-image-2 released April 21, 2026. OpenAI’s latest image model reasons before it renders — planning the layout and checking its own work — with up to 4K resolution and ~99% character-level text accuracy in English (90%+ across CJK, Hindi, Bengali, and Arabic scripts per third-party benchmarking; see coverage). It replaces both DALL-E 3 and the interim gpt-image-1.5 in ChatGPT. MCP servers are already adding support — lansespirit/image-gen-mcp was first.

spartanz51/imagegen-mcp (38 stars, up from 33) remains the most capable dedicated DALL-E server, supporting three model generations (DALL-E 2, DALL-E 3, gpt-image-1) with both text-to-image and image-to-image workflows. naporin0624/gpt-image-1-mcp has added gpt-image-2 support alongside its original gpt-image-1 batch editing, native transparency, and advanced text rendering features. SureScaleAI/openai-gpt-image-mcp (109 stars) turns out to be the most-starred dedicated OpenAI image server in this table — it had no star data at our last audit.

The gpt-image-2 release is the biggest story in image generation MCP since our initial review. With ~99% text accuracy and 4K output, it substantially reduces text rendering as a weakness of OpenAI’s models. lansespirit/image-gen-mcp (in the Multi-Provider section below) already supports it alongside gpt-image-1.5. jacwu’s Azure bridge remains notable for enterprise users routing through Azure OpenAI endpoints.

Flux (6 servers)

Flux (from Black Forest Labs, the team behind Stable Diffusion) has rapidly built an MCP presence with six servers:

jmanhype/mcp-flux-studio (25 stars, up from 24) is the most feature-complete Flux server with four tools covering text-to-image, image-to-image, inpainting, and structural controls (canny edge, depth map, pose estimation). It supports Flux 1.1 Pro, Ultra, and Dev variants. awkoy/replicate-flux-mcp (105 stars) turns out to be the most-starred server in this table (ahead of the now-archived GongRzhe server) — it had no star data at our last audit.

tehw0lf/flux-mcp stands out for local GPU execution — it runs Flux models directly on your hardware with smart VRAM management and auto-unload. Requires 12GB+ GPU VRAM but eliminates API costs entirely. It supports both FLUX.1-dev and the newer FLUX.2-dev models.

Most other Flux servers route through Replicate’s API, making them easy to set up but subject to per-generation costs.

Google Gemini / Imagen (3+ servers)

Google’s Gemini image generation capabilities continue to grow rapidly.

shinpr/mcp-image (152 stars, up sharply from 105) is the most-starred dedicated image generation server outside ComfyUI and has seen significant model upgrades. It now uses Gemini 3.1 Flash Image for fast/balanced tiers and Gemini 3 Pro Image for quality tier, with prompt optimization powered by Gemini 2.5 Flash. New features include multi-image blending for composite scenes and Google Search grounding for factually-accurate image generation. Character consistency across multiple generations remains a key differentiator.

lansespirit/image-gen-mcp (68 stars, up from 56) has expanded dramatically to become the most comprehensive multi-provider server. It now supports five OpenAI models (gpt-image-2, gpt-image-1.5, gpt-image-1, DALL-E 3, DALL-E 2) and four Google models (Imagen-4, Imagen-4 Ultra, Imagen-4 Fast, Imagen-3). (Correction, 2026-08-15 audit: previously said “six OpenAI models” while listing only five — arithmetic error, corrected.) Being the first MCP server to support gpt-image-2 gives it a notable edge. Redis caching integration is new for production deployments.

Replicate (4 servers, including the official one)

Replicate’s value proposition for MCP is access to hundreds of image models through a single API. The Replicate Official MCP server (hosted at mcp.replicate.com) is now the recommended option — it supports all operations in Replicate’s HTTP API and features auto-discovery via the MCP Registry (Replicate’s own changelog dates this Feb 10, 2026), meaning MCP clients can find and connect to it automatically via a .well-known/mcp/server.json endpoint. It can be used as a remote hosted server or installed locally via npm. deepfates/mcp-replicate (95 stars) is now formally archived (as of May 18, 2026) — the maintainer’s own archive note says: “This repo was an experiment in MCP tooling for Replicate. The company now offers an official MCP server.” gomcpgo/replicate_image_ai is interesting as a single Go binary wrapping multiple model families with image-specific features.

fal.ai (3+ servers)

raveenb/fal-mcp-server (52 stars, up from 38) provides the broadest model access of any image generation MCP server — its own README claims 600+ models spanning images, video, and audio. Its cost tracking feature lets agents monitor generation costs in real-time, which is critical for production use. Dynamic model discovery means new models on fal.ai become available automatically.

am0y/mcp-fal dropped sharply to 2 stars (from 77 at our last audit) — the repository itself shows no sign of being deleted, transferred, or recreated, so this reflects organic star churn rather than an error in this page. It still focuses on developer experience with model search/discovery tools, asynchronous execution for long-running generations, and queue management for batch workflows.

Midjourney (3 servers)

Midjourney has no official API and no official MCP server for general developers — though it has opened an Enterprise API application process (see note below). All three implementations in this table use unofficial API proxies (GPTNB, AceDataCloud) that relay requests to Midjourney’s Discord-based infrastructure. This means they can break without notice if Midjourney changes their systems, and their terms-of-service compliance is ambiguous.

z23cc/midjourney-mcp (12 stars) offers the most complete feature set with 7 tools covering generation, upscaling, variations, zoom, pan, inpainting, and face-swap. AceDataCloud/MidjourneyMCP (7 stars; repo was renamed from MCPMidjourney) goes further with roughly 14 tools including video generation and translation, but routes through AceDataCloud’s paid proxy API.

If you need Midjourney-quality output with a stable API, consider Flux or Ideogram instead — both offer similar aesthetic quality with first-party API access.

On the “enterprise API”: Midjourney’s own Enterprise API page (dated July 16, 2025) describes an application/survey process to gauge interest and gather requirements — not a public launch. (Correction, 2026-08-15 audit: earlier versions of this page said Midjourney “launched an enterprise API in late 2025.” That overstates what the primary source shows — as of this audit it remains an application/waitlist process with no confirmed public release date, so the wording has been corrected.)

Together AI (3+ servers)

Together AI servers provide a cost-effective alternative for Flux model access. Together AI’s pricing tends to be lower than Replicate for the same Flux models. The servers are simple — mostly single-tool wrappers for text-to-image generation — but they work.

Ideogram (2+ servers)

Ideogram is known for superior text rendering in generated images — a weakness of most other models. flowluap/ideogram-mcp-server offers the most complete integration with four tools covering generation, editing (with mask support), image description, and download. Ideogram v3 is the current model version.

Multi-Provider / Platform-Agnostic (4+ servers)

apinetwork/piapi-mcp-server (73 stars, up from 70) has expanded its service coverage to include Hunyuan, Skyreels, and Wan video generation plus MMAudio music creation, alongside the existing Midjourney, Flux, Kling, Luma, Suno, and Trellis integrations. This is the closest thing to a “universal creative AI” MCP server.

writingmate/imagegen-mcp (10 stars) supports SeedDream-4 (ByteDance’s image model) and Gemini 2.5 Flash Image Preview alongside OpenAI, Imagen 4, and Flux — useful for comparing outputs across six providers without switching MCP configurations.

maheshcr/image-gen-mcp includes budget alerts and cost tracking with Cloudflare R2 storage — a practical feature for teams worried about runaway API costs.

RamboRogers/cyberimage was archived on March 15, 2026 (39 stars at time of archival). Its cyberpunk-themed web UI and text-to-video features are no longer being developed.

Other Notable Servers

Adobe Firefly (NEW): msabramo/python-firefly (2 stars, Python) — the first Adobe Firefly MCP server, closing a gap flagged in our initial review. Provides Python client, CLI tool, and MCP server for Adobe Firefly API image generation. Supports aspect ratio, style, negative prompts, and seeds. Requires Adobe client credentials. Early adoption (2 stars) but functional — this is the only way to access Firefly through MCP.

Draw Things (NEW): james-see/mcp-drawthings (21 stars, up from 9; TypeScript, MIT) — local Mac image generation via the Draw Things app on Apple Silicon (M1/M2/M3/M4). Four tools: check_status, get_config, generate_image, transform_image. Uses whatever models are loaded in Draw Things — a nice local-first option for Mac users who want zero API costs without the complexity of ComfyUI.

Leonardo AI: ish-joshi/leonardo-mcp-server (4 stars, up from 2; Python) — the only Leonardo AI MCP server, with HTTP + stdio modes, job creation, model listing, and generation history. Leonardo’s strength is its fine-tuned models for game art and concept design.

Hugging Face: nikolausm/huggingface-mcp-server (1 star, up from 0; JavaScript, MIT) — wraps Hugging Face’s inference API for multiple Stable Diffusion variants with free tier access.

What’s Missing

• Adobe Firefly — Partially addressed. msabramo/python-firefly now provides MCP access to Adobe Firefly API, though adoption is very early (2 stars). Full feature coverage of Firefly’s capabilities (generative fill, text effects, etc.) is not yet available.
• Canva AI image generation — Canva’s text-to-image features still have no MCP integration.
• Dedicated inpainting/outpainting — Outside ComfyUI, few servers offer robust inpainting or outpainting tools. jmanhype/mcp-flux-studio has inpainting but the category remains thin.
• Prompt engineering helpers — No MCP server focuses specifically on optimizing prompts for different models. shinpr/mcp-image’s auto-optimization (now via Gemini 2.5 Flash) is the closest, and it’s quite good.
• Image-to-3D pipelines — Despite 3D generation models existing (TripoSR, InstantMesh), no dedicated image-to-3D MCP server exists. PiAPI’s Trellis integration is the only option.
• Consistent character/style systems — Only shinpr/mcp-image offers character consistency features. No server provides style transfer across generations in a systematic way.
• Cost comparison tools — With so many paid API options, no MCP server helps agents choose the cheapest provider for a given generation task (maheshcr’s budget tracking is per-provider, not cross-provider).
• gpt-image-2 MCP coverage — lansespirit/image-gen-mcp and naporin0624/gpt-image-1-mcp are the dedicated servers we’ve found supporting gpt-image-2 so far; broader MCP-ecosystem adoption is still catching up to the April 2026 model release.

The Bottom Line

The image generation MCP ecosystem continues to mature. 45+ servers, 12 subcategories, and coverage of every major generation platform. The April 2026 gpt-image-2 release — with near-perfect text rendering and up to 4K output — shifts the quality landscape significantly. Adobe Firefly finally has an MCP server (early stage). Three servers were archived since our last audit (GongRzhe Flux, CyberImage, and now deepfates/mcp-replicate), but new additions and organic growth (Draw Things, naporin0624 gpt-image-1, Adobe Firefly) more than compensate.

For local generation: ComfyUI servers dominate. joenorton/comfyui-mcp-server (398 stars, v1.0 rewrite) gives you the most flexibility. NEW: james-see/mcp-drawthings offers a simpler local option for Mac users via Draw Things.

For API-based generation: tadasant/mcp-server-stability-ai (now with SD 3.5) is the most polished single-platform server. For multi-model access, raveenb/fal-mcp-server’s 600+ model catalog is hard to beat. lansespirit/image-gen-mcp is the first to support gpt-image-2.

For cost-conscious users: Together AI servers offer competitive Flux pricing. tehw0lf/flux-mcp runs locally with zero API costs. maheshcr/image-gen-mcp includes budget tracking.

For quality-first users: gpt-image-2 is the new benchmark for text accuracy and resolution. Ideogram excels at text rendering today. shinpr/mcp-image (152 stars) offers the best prompt optimization. Flux 1.1 Pro remains a strong balance of quality and API stability.

The category holds at 4.0 out of 5. The gpt-image-2 release is transformative for OpenAI’s image quality. Adobe Firefly’s MCP gap is closing but adoption is minimal. Midjourney has opened an Enterprise API application process, but public developer access remains restricted — see the correction note above. The core use case — agents that can generate and manipulate images — is well-served across price points and deployment models, with the ecosystem showing continued growth (ComfyUI’s joenorton server +35% since our last audit, shinpr/mcp-image +45%).

Category: Design & Creative MCP Servers

This review was refreshed on 2026-08-15 using Claude Sonnet 5 (Anthropic). Originally published 2026-03-16.