SMF·RECRAFT
Can a prompt replace Recraft?
AI image generation — AI image and video generation
Exhibit tracking slip
Verdict
A local ComfyUI setup with a vector-friendly diffusion model can generate icon and illustration sets, and a prompt library on top of it keeps runs organized and reproducible. What it cannot do is Recraft's actual differentiator: a model specifically trained to output clean, editable vector paths rather than a raster image traced into vectors, plus a matching style system that keeps a whole icon set visually consistent.
Exhibit A — The prompt
Received on31.07.2026Build a local image-generation workflow manager: a prompt library and job queue in front of a self-hosted ComfyUI instance, for producing and organizing sets of branded assets. Do not build a new diffusion model or attempt to match Recraft's proprietary vector-output model — this wraps whatever open model the user installs in ComfyUI.
Core loop: the user writes a prompt with a negative prompt, seed, dimensions, and a chosen ComfyUI workflow (an installed model + settings preset), submits it to a local ComfyUI server, and the job queue polls for completion. Every output is saved to disk with a sidecar JSON file recording the exact prompt, seed, and workflow used, so any result can be reproduced exactly. A contact-sheet view lets the user browse, favorite, tag, and re-run past generations, and supports local image-to-image (feeding a saved output back in as a reference) without uploading anything anywhere.
This needs ComfyUI installed and running locally — either on a machine with a capable GPU, or pointed at a self-managed remote GPU box the user rents. No cloud API key is required if the user supplies their own model files with an appropriate license; without a GPU, generation is simply slow or impractical, and the README should say so rather than silently degrading.
Do not build: a new frontier diffusion model, a hosted generation service other people can use, or a claim that outputs are true vector paths — ComfyUI's standard workflows produce raster images. If the user wants an SVG, the honest scope is tracing a raster output afterward with a separate tool, not the model itself.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 a model trained specifically for clean, editable vector output rather than raster-to-vector tracing
- B.2 Recraft's style-consistency system across a whole generated set
- B.3 hosted GPU capacity, so generation doesn't queue behind your own hardware
- B.4 a licensed, curated training dataset behind the model
- B.5 mobile and web access without a machine running locally
Prior art
Exhibit C — Why people still pay: frontier models, compute, and data
People pay for Recraft because generating a raster image is the easy part — training a model that reliably outputs clean vector paths instead of a traced bitmap, and keeps a whole set of icons visually consistent, is the actual product. A local ComfyUI workflow generates raster images; getting genuinely vector output at that quality needs the model Recraft built, not a prompt box in front of a general-purpose one.
Questions
Can I import my existing Recraft projects?
No — Recraft's generation history and style presets live on their servers and aren't exportable in a form this build can read. You'd start a fresh prompt library from your first local generation.
Will it work on my phone?
Only as a client to a machine running ComfyUI elsewhere on your network. The generation itself needs a GPU, which phones don't have; you're browsing and queuing jobs, not rendering on-device.
What does it cost to run?
Nothing per generation if you own the GPU — electricity aside. Renting cloud GPU time instead runs roughly $0.50–$2 per hour depending on the card, billed only while a job is actually running.
What's the one thing that doesn't survive the rebuild?
True vector output. Recraft's model was trained to produce clean, editable paths; a standard local diffusion model produces a raster image, and turning that into a vector afterward with a tracing tool never looks as clean as output from a model built for it.
Related tools
Receipt