.ai

the best open source ai models

Guides

Rankings

FAQ

Upload

PirateDiffusion: run any model from Telegram, no GPU needed

Q: Does it work with Discord?

Through BitVector, a separate team's Discord bot and web app that run on the same hosted models. See the BitVector guide on this site.

Topic: Cloud services

By Captain

Published 2026-10-03

Latest information

This guide is a snapshot. The source below is kept up to date by its makers; check it for the newest details.

PirateDiffusion command cheat sheet

Overview

PirateDiffusion is a Telegram bot (and a companion web interface) that runs image and video models on a hosted cloud. Nothing is installed: you type a command into the chat, the picture or video comes back in the same chat a few seconds later, and every model in the ModelTrends.ai catalog that carries a Run box is one trigger word away. It is the simplest way to try a

checkpoint

LoRA

before downloading gigabytes of files, and it works the same on a phone as on a desktop.

There are three ways to create. /render is the manual, fully controlled command for

SD 1.5

and SDXL-family checkpoints, LoRAs and embeddings, with every parameter spelled out in the prompt. /workflow (short form /wf) runs hosted

ComfyUI

workflows, which is how the newer architectures are served:

video, music and editing models. Skills, written as a double slash, sit on top of both: you describe what you want and a language model picks the workflow, expands the prompt and fills in the parameters for you.

This guide is a condensed, model-centred version of the official cheat sheet: the commands you need to run a model from this catalog, the syntax that changes results most, and the fixes for the errors people hit first. The cheat sheet linked above is the complete reference and is kept current by the PirateDiffusion team.

Reference

Name	Type	What it is
/render <trigger> prompt	command	Creates images with a checkpoint. The trigger word in angle brackets picks the model, for example <realvis4-xl>. LoRAs and embeddings use the same brackets, with an optional weight: <lora-name:0.8>.
/wf /run:name prompt	command	Runs a hosted ComfyUI workflow by name. The modern models (Flux, Qwen, Wan, LTX, music) are served this way. /workflow /show:name prints the workflow's own help text and image fields.
// prompt	skill	Lets the Skills system choose the best workflow and expand your prompt. /skill /list shows what exists; /skill /run:name runs one directly.
(( )) and [[ ]]	weighting	Round brackets emphasise, square brackets de-emphasise. Each nested pair is a 1.1x step. Never nest one kind inside the other.
#recipe	recipe	A community prompt template that adds quality tags, negatives, sampler and sometimes a model. #boost, #sdxlreal, #nfix and #everythingbad are popular. One recipe per prompt.
/concepts, /concept /search:word	lookup	Lists trigger words, searches models, and /concept /recent recalls what you tried. /concept /default:name sets your default checkpoint.
/size:WxH, /portrait, /landscape, /wide, /tall	size	Output size. SD 1.5 is trained at 512, SDXL at 1024; going far above a model's native size repeats heads and limbs.
/guidance:N, /steps:N, /seed:N, /images:N	sampling	CFG (about 7 for most checkpoints, 1.5 to 4 for LCM), step count (25 is average), a fixed seed for repeatable results, and how many images to make.
/sampler:name /karras	sampler	/samplers lists them. dpm2m with /karras is the usual recommendation; lcm trades quality for speed at very low steps.
/adetailer	detailer	Finds faces, eyes and hands after the render and redraws them. Works on SD 1.5 and SDXL.
/highdef, /facelift, /more	reply commands	Reply to an image: /highdef doubles the pixels, /facelift upscales 4x with optional /photo or /anime mode, /more makes variations and accepts /strength, /guidance and /size.
/remix, /inpaint, /outpaint, /bg	reply commands	Image-to-image with a new prompt and model, mask-and-repaint in the browser, canvas expansion, and background removal.
/showprompt, /describe, /download	reply commands	Shows how an image was made, writes a prompt for a non-AI photo, and fetches the uncompressed original instead of Telegram's compressed copy.
/vae:name, /nofix, /clipskip:2, /lpw, /parser:new	advanced	VAE override for washed-out colour, SDXL refiner off, clip skip, long-prompt mode, and per-token weights such as (blue cat:0.4).

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Open the PirateDiffusion bot in Telegram, type /start, and then /debug to confirm the account is activated. Activation is explained on the PirateDiffusion site.
Find a model: type /concepts, or /concept /search:realistic, or open a model on ModelTrends.ai and read its trigger word from the Run box.
Make a first image: /render a grey tabby cat in a garden <realvis4-xl> #boost. The echoed confirmation tells you how many images and how long.
Tune the prompt: put the important words first, add (( )) around what matters and [[ ]] around what must not appear, and keep one recipe.
Fix faces and hands: add /adetailer to the render, then reply to the best result with /highdef and afterwards /facelift.
Run a modern model: /workflow /show:name to read the fields, then /wf /run:name your prompt. For image input, reply to the image or pass /image1:ID as the workflow describes.
Let skills do the work: type // a warrior standing in the rain, or // edit, make her hair blonde while replying to a photo. Vote on results so the system learns your taste.
Save your setup: /settings /save:my-anime-set stores model, sampler and steps as a loadout; /settings /load:my-anime-set brings it back.

Examples

A balanced SDXL render

/render /size:1024x1024 /sampler:dpm2m /karras /guidance:7 /steps:30 <realvis4-xl> [[<fastnegative-xl:-2>]] ((high quality, masterpiece)) a portrait of a fisherman mending nets at dawn, harbour fog [[blurry, low resolution, deformed, extra fingers]] /adetailer

Fast LCM drafts

/render /sampler:lcm /guidance:1.5 /steps:6 /images:9 /size:1024x1024 <realvis2-xl> /nofix oil painting of a wizard in a desert, retro fantasy art

A hosted workflow with an input image

/workflow /show:edit-qwen

/wf /run:edit-qwen /initimage:Ixyz123 change the shirt to red

Skills

// a warrior standing in the rain, cinematic

// edit, make her hair blonde

/skill /run:anime-illustrious a shrine maiden under falling leaves

Reply commands

/highdef

/facelift /photo

/more /strength:0.5 /size:1536x1536

/remix a watercolor painting <sdxl>

/showprompt

Tips

One checkpoint per prompt. If you name two base models, the last one wins; one to three LoRAs on top is the sweet spot.
SD 1.5 and SDXL do not mix. The bot refuses to render a 1.5 LoRA on an SDXL checkpoint, which is a feature.
Negative embeddings go in double square brackets with a negative weight: [[<fastnegative-xl:-2>]]. They are the biggest single quality boost for SDXL renders.
Steps presets: /steps:wayless gives nine fast drafts, /steps:waymore gives two slow high-quality images. Draft cheap, finish expensive.
Telegram compresses images. Reply with /download for the original file, and add /format:png to a render when you need lossless output.
Use /translate to prompt in your own language; the bot detects it. Avoid slang.
/ping tells you whether the queue, the bot or Telegram is the problem. /start nudges a sleeping bot.
Skills accept a language model override; the default picks for you, but a skill's own template can name the LLM it wants.

Include the word edit for edits, start video requests with a video of, downvote the result, or run the skill you want directly with /skill /run:name.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.

Try PirateDiffusion

Questions

Do I need a GPU or any software?

No. Everything runs on the hosted cloud. Telegram on a phone is enough; the web companion opened with /webui adds file management and the

inpainting

canvas.

How do I run a model from this catalog?

The Run box on a model page carries the trigger word. Checkpoints, LoRAs and embeddings go in angle brackets inside /render; newer architectures run through a named workflow with /wf /run:name.

What is the difference between /render and /workflow?

/render drives SD 1.5 and SDXL-family checkpoints directly with every parameter in the prompt. /workflow runs a hosted ComfyUI graph, which is how Flux, Qwen, Wan, LTX and audio models are served. Skills choose between them for you.

Can I upload my own ComfyUI workflow?

Yes. Export the workflow and its API-format JSON from ComfyUI with dev mode on, upload both in the Workflows manager, map the fields (seed, guidance, size, sampler, steps, prompts) and the models, then run it with /wf /run:name.

Is my chat private?

A one-to-one chat with the bot is private by default. In public groups, /history and /skills /list only show your own items, but anything you render in a group is visible to that group.