Visual DSL for ComfyUI. Pure Lua. Zero runtime dependencies for DSL and compilation. Server execution (connect/run) requires curl or a custom HTTP backend.

Transforms semantic scene composition into ComfyUI node graphs. Images become portable project files through PNG-embedded recipes.

luarocks install vdsl

For pod-side orchestration (Profile DSL apply, model sync, batch generation), vdsl-mcp is required as the companion MCP server. The Lua side emits canonical manifests; vdsl-mcp expands them into MCP tool calls (vdsl_profile_apply, vdsl_project_init, vdsl_run_script, …) on the client side.

local vdsl = require("vdsl")

local w = vdsl.world {
  model     = "sd_xl_base_1.0.safetensors",
  clip_skip = 2,
}

local hero = vdsl.cast {
  subject  = "warrior woman, silver armor, dynamic pose",
  negative = "blurry, low quality, deformed",
  lora = {
    { name = "add_detail.safetensors", weight = 0.7 },
  },
}

local result = vdsl.render {
  world = w,
  cast  = { hero },
  steps = 30,
  cfg   = 7.5,
  seed  = 42,
  size  = { 1024, 1024 },
}

print(result.json)

All entities are immutable. Every mutation returns a new instance.

• Compile -- vdsl.render() produces a ComfyUI-compatible prompt JSON
• Themes -- Built-in presets (cinema, anime, architecture) with lazy loading
• Embed -- vdsl.embed() writes recipe + prompt into PNG tEXt chunks
• Import -- vdsl.import_png() reads back vdsl recipes or decodes ComfyUI prompts
• Registry -- vdsl.connect() queues jobs to a running ComfyUI instance
• Run -- vdsl.run() full pipeline: compile, queue, poll, download, embed
• Serialize -- entity:serialize() projects any entity to a plain, JSON-ready table; Entity.deserialize(type, t) restores it losslessly. Identity reuse and versioning are plain serialize -> store -> load: the application owns the storage backend and version history (see docs/entity-architecture.md)
• Cam -- vdsl.cam{ world, base, shots, negative?, common? } renders a sequence of shots through one identity Subject (each shot overlays a per-shot trait + an optional shared common trait), emitting one image per shot (see docs/cam-and-identity.md)
• Over-prompt lint -- vdsl.check emits warn-only diagnostics when a single prompt category overflows (lighting / 2D-style words), naming the offending clauses; Shot:intent(category) silences a deliberate one (see docs/over-prompt-lint.md)
• Profile -- vdsl.profile{} declares the whole pod (ComfyUI / vLLM / Ollama, models, custom nodes, sync routes) as a canonical JSON manifest

vdsl.profile{} describes a reproducible ComfyUI / vLLM / Ollama pod declaratively: ComfyUI ref, Python deps, custom nodes, models, B2 sync routes, env, and install hooks. Output is a canonical JSON manifest (SHA256 = profile_hash).

local vdsl = require("vdsl")

local profile = vdsl.profile {
  name = "fantasy",

  comfyui = { ref = "v0.3.26" },        -- optional: omit for vLLM/Ollama-only pods
  python  = { version = "3.12", deps = { "xformers==0.0.27" } },

  custom_nodes = {
    { repo = "ltdrdata/ComfyUI-Manager", ref = "main" },
  },

  models = {
    { kind = "checkpoint",
      dst  = "sd_xl_base_1.0.safetensors",
      src  = "b2://vdsl-assets/checkpoints/sd_xl_base_1.0.safetensors" },
  },

  sync = {
    push = { "/workspace/ComfyUI/output/ → b2://vdsl-output/{pod_id}/" },
  },
}

print(profile:manifest_json(true))

Convergence is orchestrator-driven — the manifest is consumed by the vdsl-mcp vdsl_profile_apply tool, which expands it into a sequence of MCP tool calls on the client side. The pod itself runs no convergence script.

Source schemes for models[].src and sync routes are limited to b2:// (Backblaze B2) and file://. Stage external assets into B2 before referencing them. B2 credentials are MCP-owned and auto-injected at apply time — never declare them in the profile.

See examples/11_profile.lua (ComfyUI), examples/12_vllm_profile.lua (vLLM serving), and docs/profile-and-orchestration.md for the full design.

See SPEC.md for the full specification, and docs/profile-and-orchestration.md for Profile / orchestration design.

MIT