> ## Documentation Index > Fetch the complete documentation index at: https://docs.seriesmith.com/llms.txt > Use this file to discover all available pages before exploring further. # Setting up seriesmith > Where shots render: your ComfyUI, Comfy Cloud, a deployment or Comfy Router, and the settings that choose. ## Install Inside ComfyUI, nothing: link the node pack (below) and seriesmith runs on ComfyUI's own Python, loaded from the repo the pack is linked from. For `seriesmith serve` and the command line, Python 3.10 or newer, in a virtual environment of its own (`run_seriesmith.bat` / `run_seriesmith.sh` make it the first time they run): python -m venv .venv .venv\Scripts\activate (Windows; source .venv/bin/activate elsewhere) pip install -e . That brings comfy-sdk, which Comfy Cloud, deployments and Router use. Inside ComfyUI, seriesmith installs comfy-sdk into ComfyUI's Python itself, in the background, once a Comfy API key is set (saved in Settings, or `COMFY_API_KEY` at startup); Settings shows how it went and has **Install** to try again. Only those backends use it. `SERIESMITH_NO_AUTO_INSTALL=1` turns the automatic install off. ## Your own GPU: ComfyUI Local renders go straight to your ComfyUI (`comfyui_url`, default `http://127.0.0.1:8188`; inside ComfyUI, that ComfyUI on whatever port it runs), through its own API: the input files go up to `input/seriesmith/` (named by their content, so each is sent once), the graph is queued on `/prompt`, followed on `/ws` and `/history`, and its outputs fetched from `/view`. A job that was running when seriesmith stopped is picked up again by its prompt id. ### Through comfy-api-proxy instead (optional) **comfy-api-proxy** serves the same Comfy API v2 as Comfy Cloud in front of a ComfyUI. seriesmith used it for every local render before it talked to ComfyUI directly; it still can, with `"local_via": "proxy"` in the settings (or `SERIESMITH_LOCAL_VIA=proxy`, or Settings → Through comfy-api-proxy). Install it in the same environment: pip install comfy-api-proxy comfy-api-proxy start --comfyui [http://127.0.0.1:8188](http://127.0.0.1:8188) --port 8189 --state-dir "%LOCALAPPDATA%\seriesmith\proxy" * `--state-dir` matters: without it, a restart of the proxy forgets the files it has uploaded and the jobs it has run. * It listens on 127.0.0.1 only. To reach it from another machine it needs `--token`, and seriesmith's `local_url` set to match. * `comfy-api-proxy status` / `stop` manage the background process. * The proxy can't see ComfyUI's model lists or installed nodes; seriesmith asks ComfyUI for those directly (`comfyui_url`). * With `local_via` "proxy" you rarely start it yourself: `seriesmith serve`, and every command that queues work on `local` (render, upscale, post, master, tour, kreagen), start it when it isn't running and ComfyUI is. Check it: seriesmith backends # each backend, and whether it answers seriesmith backends --try local # a picture up through ComfyUI and back ## The seriesmith nodes Everything seriesmith sends a backend runs on its own nodes (`comfy_nodes/`): the shot loader and saver, the `continuous: latent` chain, the upscale loaders and saver, and the upscale and post compute nodes. h3pipe's node pack isn't needed (the runner refuses a graph that would use it). Link the folder into ComfyUI's `custom_nodes` and restart ComfyUI: mklink /J "\\custom\_nodes\seriesmith" "\\comfy\_nodes" (Windows) ln -s "\/comfy\_nodes" "\/custom\_nodes/seriesmith" (elsewhere) They sit happily beside h3pipe's node pack if you keep it for h3pipe. The pack also serves the editor inside ComfyUI, the same app as `seriesmith serve` (INSTALL.md, "The editor inside ComfyUI"): it finds the seriesmith package in the repo it is linked from (or installed in ComfyUI's Python, if it is). No other pack is needed to render. A node a target can do without (binding `bypass_if_missing`: H3's `H3SLAAttention`, from ComfyUI-PlagueKind-Nodes) is spliced out of the graph where ComfyUI hasn't it, and the take notes it. `seriesmith render --dry-run` writes the graph a shot would run. ## The default models A series config that names no video target (`series.target`) or picture target (`refs.target`) renders on this machine's default for that kind: 1. the one picked in Settings (`default_targets` in the settings file); 2. else the one found installed (`detected_targets`). Each time a host (`seriesmith serve`, the editor inside ComfyUI) sees ComfyUI answer, and on each readiness check after that, it takes the first target, in `default_rank` order (target.json), that is ready, else the first that is degraded: H3 Ref2VA, LTX-2.5, LTX-2.3 ingredients, Wan 2.2 VACE / I2V / 5B; Krea 2, FLUX.2 Klein, Z-Image Turbo, Qwen-Image 2.1, FLUX.3 Klein, MOSS-Image SFT, ...; 3. else the built-in: `minimax_h3_ref2va` and `krea2`. A new series (the starter, a new series from the wizard, a Simple-mode project) gets the defaults written into its series config when they aren't the built-in ones, so it renders on the same models wherever it is opened. Hosted (Router) models are never picked by themselves: they are paid. ## Comfy Cloud, deployments and Router All three use your Comfy API key (platform.comfy.org). Keep it in the environment rather than in a file: setx COMFY\_API\_KEY comfyui-... (Windows; export COMFY\_API\_KEY=... elsewhere) Router also needs it when everything else runs locally: a hosted model fetches its reference pictures from Comfy Cloud, so they are uploaded there (a few models take them inline in the request instead). The editor's **Settings** window sets the key too (saved in the settings file below); `seriesmith backends` says whether one is set. ### Comfy Cloud Comfy Cloud is ComfyUI on Comfy's GPUs, on your Comfy subscription: the open-weights models (MiniMax H3, LTX-2, Wan 2.2, FLUX, ...) with no GPU of your own and nothing to download. Pick `cloud` as a shot's, an episode's or the series' backend (or `seriesmith render ... --backend cloud`). Cloud runs ComfyUI's own nodes and the packs Comfy has installed there, not seriesmith's, so a shot sent there is made of Cloud's nodes alone: seriesmith cuts the reference pictures and voice slices itself and sends them as plain uploads, and Cloud's own save nodes return the video and its sound. What comes back is a take like any other (thumbnail, strip, record). Model files are Cloud's: a model whose accelerator LoRA Cloud hasn't renders its base preset there (H3: 20 steps, no LoRA, slower), and a node a model can do without (H3's sparse attention) is left out; the take's notes say so. What stays on your own ComfyUI (or a deployment): a `continuous: latent` shot, which holds the take before it through seriesmith's chain nodes. A take made on Cloud keeps no latent (an upscale of it goes through the VAE), no PNG frames and no foley bed. ### Comfy Router models A hosted model is a target like any other, on the `router` backend only: no GPU, no model files. seriesmith ships the video and image models it has tried (Seedance, Kling, Wan, Veo, Runway, hosted H3; Nano Banana, Seedream, GPT Image, Flux, Qwen, Krea 2 and more), and others as **drafts**, which stay out of the pickers until you probe one and **Enable** it in the editor's **Models** window. That window also adds any other model from Router's catalog ([EDITOR.md](/editor), "Comfy Router models"). What a hosted take costs is the provider's, billed to the key's account. Runs on Cloud, Router or a deployment are counted in the Jobs window, and `paid_limit` (below) caps them per server session. ## Settings Anything not in the environment can go in a settings file, `%APPDATA%\seriesmith\settings.json` (Windows) or `~/.config/seriesmith/settings.json`: ```json theme={null} { "comfyui_url": "http://127.0.0.1:8188", "default_targets": {"video": "ltx2"}, "deployments": {"minimax_h3_ref2va": "https://.run.comfy.app"} } ``` The other keys, most of them written by the editor rather than by hand: * `cloud_url` (default `https://cloud.comfy.org`) and `api_key` (better kept in `COMFY_API_KEY`). * `default_targets`: `{"video": id, "image": id}`, the models a series naming none renders on (Settings' **Video model** / **Picture model**); `detected_targets`: the ones found installed, written by seriesmith (**The default models**, above). * `hf_token`: a Hugging Face read token, for the gated downloads (LTX-2.5, FLUX) the Features window makes; `HF_TOKEN` in the environment wins. * `ffmpeg_dir`: a folder holding your own ffmpeg and ffprobe, used before anything else (`seriesmith/ff.py`). * `comfyui_dir`: ComfyUI's folder (the one with `main.py`), for the Features window to install into when the standalone server can't ask ComfyUI (`/internal/folder_paths`). * `local_via`: `"comfyui"` (the default: local renders straight to ComfyUI) or `"proxy"` (through comfy-api-proxy at `local_url`, default `http://127.0.0.1:8189`). * `paid_limit`: how many paid runs one server session may queue (none: no cap). * `enabled_drafts`: the built-in draft models switched on for every show (the Models window's **Enable**). * `target_limits`: this machine's changes to a built-in model's limits (its **Limits** form; **Back to shipped** removes one). Beside the settings file live the models you saved for every show (`targets//`), the editor's project folders (`default/seriesmith/config.json`) and the queue's journal (`queue.json`; `queue-comfyui.json` for the editor inside ComfyUI). A model saved for one show only goes in `/targets//` instead. The environment wins over the file: `SERIESMITH_LOCAL_URL`, `SERIESMITH_COMFYUI_URL`, `SERIESMITH_CLOUD_URL`, `SERIESMITH_LOCAL_VIA`, `COMFY_API_KEY`, `HF_TOKEN`; `SERIESMITH_SETTINGS` names another file. These are this machine's settings; nothing here goes in the series config, which gets shared. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.