Skip to main content
There are two ways to add a video or image model to seriesmith:
  • In the app. The Models window turns a Comfy Router model, or a workflow you’ve saved in ComfyUI, into a model your shows can use. No code. It’s kept on your machine, for every show or for one show, and you try it before you switch it on.
  • In seriesmith’s code. A model that ships with seriesmith, for everyone: a folder under seriesmith/targets/, the files generated from it, and its tests. This is for contributors.
Either way a model is the same thing underneath: a target.json that says what the model takes (pictures, length, size, sound) and where each of a shot’s values goes. The app writes that file for you; in the code you write it, or generate it.

​
In the app

​
The Models window

Everything about the models a show can use is in one window. Open it from the Shots tab, Model: → Add / validate models…, or from the sliders button beside any model picker. It opens on All targets:
  • Add a model: a Comfy Router model from its catalog, or a workflow of yours.
  • Hosted models to try: the Router models seriesmith ships untested (drafts).
  • Your own models: the ones added here, for every show or this show only.
  • Each model’s details say what files it needs, where its graph comes from, and carry its Limits.

​
A Comfy Router model

A hosted model needs your Comfy API key (Settings) and no GPU. Choose Add a Comfy Router model…:
  1. Pick it from Router’s catalog. Type any part of its id, and narrow it to Video, Image or All. A model you’ve already added says so. One whose name says text-to-video is marked: seriesmith draws from pictures, so it will be refused.
  2. Read it. seriesmith reads the model’s published schema and proposes where a shot’s (or a ref’s) prompt, pictures, length, size, seed and sound go. A field written for several models at once is flagged Check this, with its text beside it. That’s where a limit hides that the schema only states in prose: Seedance 2.0’s 4–15 seconds, Wan 2.7’s 720P/1080P.
  3. Correct its limits in the form below it: what the prose said.
  4. Dry run (free). Pick a shot (or a ref) and see the exact request, with nothing sent.
  5. Save it for every show or for this show only. It’s saved as a draft.

​
A model from your own workflow

Build the graph in ComfyUI and save it, then choose Add a model from a workflow…:
  1. Pick the workflow, say what it makes or leave it on auto, and Read it. Auto reads the graph: a video node, a frame count or a video save makes it a video model; otherwise it’s an image model. seriesmith works out which widget takes the prompt, the negative, the size, the length, the seed, the steps, the sampler and each model file.
  2. It can end in any save node: SaveVideo, SaveWEBM, Video Combine (with its sound), a frames saver, SaveImage, a node pack’s own. The graph’s best output is the take, and seriesmith saves it in that node’s place.
  3. An image model’s LoadImage nodes are its pictures. Without any it’s text to image; with them it edits, and each take’s pictures go into them in order (an edit’s source, or a keyframe’s characters and then its plate).
  4. Read the notes. Errors stop it being saved: the graph itself is wrong. Warnings are worth checking: a length that is a guess, where a keyframe goes in, a model file seriesmith can’t check. Notes say how the graph was read.
  5. Answer what no graph can state (for a video model: its frame grid, size multiple, fps, whether it makes sound, whether dialogue is spoken) and Save, for every show or this show only. It’s a draft.
The same without the app: seriesmith target-from-workflow <workflow> [<episode>] --save (add --every-show for every show).

​
Try it, then switch it on

A model you add, and every untested model seriesmith ships, starts as a draft: it isn’t in the pickers yet.
  1. Probe it from the Models window: one take, on a shot for a video model or a ref for an image model (a character’s first view, or a keyframe to see its pictures go in). A Router probe is a paid call; a local one runs on your ComfyUI.
  2. Look at the take. Is it the right length and shape? Do the pictures you sent show in it? Is the sound there, if it should be?
  3. Enable it. It joins the pickers in every show. Back to draft takes it out again.
Limits are in each model’s details: its lengths, sizes, the longest prompt, how many pictures and how they’re sent, whether it makes sound. Edit changes them. On a model seriesmith ships, the change is kept on this machine (Back to shipped undoes it); on your own model it’s written into its target.json.

​
Where they’re kept

Which drafts you’ve enabled is kept in seriesmith’s settings (enabled_drafts). seriesmith targets lists every model and whether the ComfyUI it talks to can run it.

​
In seriesmith’s code

A model that ships with seriesmith is a folder, seriesmith/targets/<kind>/<id>/, where <kind> is video, image or audio, and <id> is the name scripts and the API use (ltx2, seedance_2_5). Its target.json has these parts: The Features window and INSTALL.md’s model list are built from models and downloads. A download link must be one recorded there: the tests fail on a link that isn’t.

​
A Comfy Router model

Generate it from its schema rather than writing it by hand, so that a change on Router’s side shows up.
  1. Save its schema: tests/fixtures/router_schemas/<provider>__<model>.json, from GET /v2/models/<provider>/<model>/openapi.json (router_inspect.fetch_schema).
  2. Add a row to MODELS in tools/make_router_targets.py: the Router id, the target id, its label and short name, its limits, whether it’s a draft, and a hook.
    • Limits are what the schema says only in prose ({"length": {"seconds": {"min": 5, "max": 20}}}).
    • The hook fixes anything else the proposer can’t read, such as aspect ratios listed in a description, or resolutions named hd / fhd.
  3. Generate: python tools/make_router_targets.py. --check fails when a target no longer matches its schema.
  4. If the proposer refuses it (“takes no pictures”), it didn’t recognise the model’s picture field. Teach router_inspect.propose that field’s name rather than writing the target by hand (that’s how FLUX 3 video’s keyframes was added).
  5. Pin its requests: python tests/test_router_golden.py --update writes tests/golden/router_calls/<id>.json, the exact request it sends for a set of shots. Read it. Keep the new file and the changes you meant: the update rewrites the others’ line endings too.

​
A local ComfyUI model

  1. Its workflow: workflow.json in the folder (API or UI format), and a binding naming the widget that takes each of a shot’s values (prompt, negative, seed, steps, size, length, model files) and the node that saves the take.
  2. Its code, only when it needs more than a value in a widget (reference pictures, keyframes, a second sampling stage): a compile.py or target.py beside it.
  3. Its files, under models: each with a family and name patterns, so seriesmith recognises an installed file and accepts a substitute of the same family. A new architecture also needs a header signature in seriesmith/targets/modelid.py, or seriesmith can’t tell its files apart by what’s in them. Each file needs a downloads entry.
  4. Its tests check its graph against a trimmed copy of ComfyUI’s /object_info (tests/fixtures/workflows/object_info_*.json). Add any new node classes to it, copied from a live ComfyUI, keeping the file’s own formatting.

​
Every new model

  • Regenerate what’s built from the targets, and commit the results:
    • python tools/make_models_md.py: INSTALL.md’s model list.
    • python tools/make_refs_md.py: AUTHORING.md’s table of what each model takes in pictures.
    • When you’ve also written about the model in AUTHORING.md: python tools/make_prompts.py (the script prompts and skill) and python tools/make_site_docs.py (this site).
  • The lists of every target in the tests:
    • tests/test_targets.py names the targets that aren’t generated from Router.
    • Some error messages in tests/golden/errors/ list every known video target. Update them with python tests/test_golden.py --update, run with the tests’ own empty settings (SERIESMITH_SETTINGS=tests/_no_settings.json). Otherwise the models on your machine end up in the files.
  • Say what it’s for: a row in AUTHORING.md’s “Choosing a video target” table (or a line in its hosted-model section), and a PLAN.md entry for what was tried live and what wasn’t.
  • Test it: .venv\Scripts\python -m pytest, then render one take live: a local model on your ComfyUI, a Router model once (paid).