PixelDojo Image Generator

The name is a lie, sort of. PixelDojo Image Generator doesn't generate anything - not locally, anyway. This node is a thin API wrapper around pixeldojo.ai's hosted Flux endpoint. You type a prompt, it phones home to their servers, and an image comes back. No model downloads, no VRAM gymnastics, no 12GB of weights parked in your models folder. If you want Flux-class quality without owning the hardware to run Flux, this is the whole point: your machine is just a keyboard and a fetch call.

That trade is real, though. This node is pay-per-image (1 credit per image, and you need an active PixelDojo subscription plus a valid API key). It's also a one-way street - you can't wire a local checkpoint into it or tweak the sampler. The sampling settings you'd normally obsess over for FLUX.1 dev (Euler, Simple scheduler, guidance 3.5) are all decided server-side. You get prompt, model, aspect ratio, seed, and out pops a finished image. For a quick test of an idea without context-switching out of ComfyUI, that's genuinely handy; for someone who owns a 4090, it's mostly a curiosity.

How it works

Under the hood it's a requests.post to https://pixeldojo.ai/api/v1/flux with your key as a Bearer token, a JSON payload of your settings, and a generous 600-second timeout. The response comes back as URLs, which the node downloads, decodes with PIL, and returns as a proper ComfyUI IMAGE tensor - so the output behaves like any other image node downstream.

The API key lookup order is: api_key_override field first, then the PIXELDOJO_API_KEY environment variable, then a pixeldojo_api_key.txt file sitting in the extension's folder. Set the env var and forget the other two.

The inputs that matter

• prompt - multiline text, freeform. Flux's T5 encoder handles natural language well, so write a real sentence rather than tag soup.
• model - four choices: flux-pro, flux-1.1-pro, flux-1.1-pro-ultra, and flux-dev-single-lora. The first three are plain text-to-image; the LoRA one is the only model that honors lora_weights and lora_scale (default 0.7, range 0–1).
• aspect_ratio - seven presets from 1:1 to 16:9, 9:16, 4:3, 3:4, 3:2, 2:3. No freeform resolution.
• num_outputs - 1 to 4 images per run.
• seed - optional, and here's the gotcha: it's only sent to the API when it's greater than 0. Seed 0 means "roll the dice," so if you set a seed and get a different image, check that you didn't set it to 0.

Output and wiring

One output, IMAGE, holding a batched tensor of everything the API returned. Feed it into a Preview Image or Save Image node - the included basic_pixeldojo.json example does exactly that - or route it into an upscaler to fix the relatively modest output resolution. Save via ComfyUI and the image lands on your disk; PixelDojo only keeps its copy of the remote file for 24 hours, so grab it before then.

Installing it

ComfyUI Manager, search "PixelDojo," hit install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/blovett80/ComfyUI-PixelDojo

Then restart ComfyUI. There's no requirements.txt - the node only imports requests, torch, PIL, and numpy, all of which ComfyUI already ships. No model files to download, no heavy deps. The only real setup is the API key from your PixelDojo account.

Where people get burned

• "API key not found" - you skipped all three key options, or the env var isn't visible to the ComfyUI process.
• LoRAs that do nothing. This is the big community complaint. The node only sends LoRA params for flux-dev-single-lora, and even then, CivitAI-hosted LoRAs have been reported as silently ignored unless you use a CivitAI key - plus LoRAs trained on quantized checkpoints just don't map onto PixelDojo's full-weight model. If a LoRA appears loaded but the output is identical with and without it, that's why.
• No images at all - check for content-policy rejections in your prompt, and confirm you have credits. The error lands in the ComfyUI console.

One last thing: PixelDojo is a real, working service, but it's a small player, and threads about it on Reddit occasionally read like marketing. The service is legitimate; the hype around it isn't always.