Discord Bot

mold includes a built-in Discord bot that connects to mold serve, allowing users to generate images and videos via slash commands.

Running

# Server + bot in one process
MOLD_DISCORD_TOKEN="your-token" mold serve --discord

# Or run the bot separately (connects to a remote server)
MOLD_HOST=http://gpu-host:7680 MOLD_DISCORD_TOKEN="your-token" mold discord

Setup

1. Create a Discord application at the Developer Portal
2. Create a bot user and copy the token
3. Invite with: https://discord.com/api/oauth2/authorize?client_id=YOUR_APP_ID&permissions=51200&scope=bot%20applications.commands (Send Messages, Attach Files, Embed Links + slash command registration)
4. No privileged intents are needed (slash commands only)

Slash Commands

Command	Description
/generate	Generate an image or video, including attachment-driven LTX-2 audio-to-video, retake, and keyframe modes and ordered MiniMax H3 references
/mesh	Generate a Hunyuan3D GLB from one source image or semantic front/left/back/right multiview attachments
/identity	Generate an image conditioned on a face reference photo (PuLID), with identity_strength and identity_start_step
/expand	Expand a short prompt into detailed generation prompts
/remix	Rewrite a prompt into subject-preserving alternatives, one creative dimension each (dimensions, style, variations 1-5)
/models	List available models with download/loaded status
/status	Show server health, queue summary, and every GPU/MIG device; large fleets paginate across limit-safe follow-up embeds
/quota	Check your remaining daily generation quota
/admin reset-quota	Reset a user's daily quota (requires Manage Server)
/admin block	Temporarily block a user from generating (requires Manage Server)
/admin unblock	Unblock a previously blocked user (requires Manage Server)

/identity

Face-identity conditioning has its own command rather than options on /generate. Discord caps a chat-input command at 25 options and /generate is already at exactly 25, and identity is qualified only for a fixed list of FLUX and SDXL checkpoints (see the Identity Photos guide), so none of /generate's video and conditioning options apply to it anyway.

Option	Purpose
prompt	Required. What to render.
identity	Required. Face reference photo as a PNG or JPEG attachment.
model	Identity-capable checkpoint. Autocompletes only models the server advertises; defaults to one.
identity_strength	0.0–3.0, default 1.0.
identity_start_step	First identity-conditioned denoise step; must be under the resolved step count. Default 0.
width / height	Output size in pixels.
steps	Inference steps.
guidance	Guidance scale.
seed	Seed for reproducibility.

Preconditions are checked in cost order, so an impossible request never takes a quota slot or a download: the declared attachment size and container and the strength range first, then the model gate against the server's advertised /api/models[].supports_identity (an absent field is read as "no", which covers both a server too old for identity conditioning and one whose binary cannot execute it) then the start step against the resolved step count, and finally the downloaded bytes. A server advertising no identity-capable model at all says so instead of guessing a checkpoint. The result embed carries an Identity row naming the photo, the strength, and the start step.

/mesh

Mesh generation has its own compact command because /generate already uses Discord's 25-option ceiling. Attach source for a single-view Hunyuan3D model, or any non-empty subset of front, left, back, and right for a 2mv checkpoint. The two forms are mutually exclusive. If model is omitted, source selects the ordinary Hunyuan3D default and named views select the five-step hunyuan3d-2mv-turbo:fp16 tier. Optional seed, texture, octree, threshold, and target_faces values ride the same generation request and admission checks as every other client. The response includes the poster and GLB using the existing Discord mesh delivery policy.

Configuration

Variable	Default	Description
MOLD_DISCORD_TOKEN	--	Bot token (required; falls back to DISCORD_TOKEN)
MOLD_HOST	http://localhost:7680	mold server URL
MOLD_DISCORD_COOLDOWN	10	Per-user cooldown (s)
MOLD_DISCORD_ALLOWED_ROLES	--	Comma-separated role names/IDs for access control (unset = all)
MOLD_DISCORD_DAILY_QUOTA	--	Max generations per user per UTC day (unset = unlimited; 0 = block all)

Video generation

Running /generate against a video model (ltx-video-*, ltx-2-*, wan*) produces an MP4 by default. Pass video_format: Animated GIF to receive a GIF instead. You can also attach a source_image for img2img on regular models, or as the first frame for LTX-2 image-to-video. When the rendered MP4 exceeds Discord's upload ceiling the bot falls back to the always-bundled GIF preview.

Use duration for the simple path: duration: 10 uses the selected model's default FPS and converts ten seconds to the nearest frame count on the selected model's advertised grid (8n+1 for the LTX families, 4n+1 for Wan). The bot uses the same advertised frame/FPS defaults as Studio, supports LTX-2's full 20-second single-generation limit, and rejects a duration beyond the selected model's limit. The existing frames and fps options remain available for precise control; duration and frames cannot be combined.

LTX-2 specialized modes are selected by their attachments: audio_file starts audio-to-video, source_video plus both retake times regenerates that time range, and two or three keyframe_* images are spaced across the requested frame count for interpolation. These modes are mutually exclusive in one command and do not need the pipeline option.

MiniMax H3's ordered references are their own pair of attachments, reference_1 and reference_2, each an image, an H.264 MP4, or a WAV. Ordering is explicit, so reference_1 must be present before reference_2. They cannot be combined with source_image, source_video, audio_file, the keyframe_* images, either retake time, or the pipeline option.

Negative prompts: leaving negative_prompt unset applies the model's advertised default negative (Wan ships a tuned one). To explicitly disable the negative prompt, pass negative_prompt: none (case-insensitive; - also accepted); Discord's 25-option cap means the opt-out rides the existing option as a sentinel rather than its own toggle.

Re-register after command-option changes

/generate's prompt option changed from required to optional so an LTX-2 image-to-video run can be submitted with just a source_image. Discord caches command definitions, so the bot's slash commands must be re-registered after upgrading before users see the optional prompt or new duration option. The new /mesh command also appears only after re-registration. The prompt relaxation is guarded on both ends: the bot only skips the up-front check when visual conditioning (source_image, source_video, or keyframes) is attached, and the server's family-aware validator still rejects an empty prompt for every image family and for unconditioned text-to-video.

Block List

The /admin block command stores blocks in memory. Blocks clear when the bot restarts. For permanent restrictions, use role-based access via MOLD_DISCORD_ALLOWED_ROLES.

NixOS

services.mold.discord = {
  enable = true;
  # tokenFile is loaded via systemd EnvironmentFile.
  # the file must contain: MOLD_DISCORD_TOKEN=your-token-here
  tokenFile = config.age.secrets.discord-token.path;
  moldHost = "http://localhost:7680";
  cooldownSeconds = 10;
};